High Resolution Forest Masking for Seasonal Monitoring with a Regionalized and Colourimetrically Assisted Chorologic Typology

Abstract

Comparisons of recent global forest products at higher resolutions that are only available annually have shown major disagreements among forested areas in highly fragmented landscapes. A holistic reductionist framework and colourimetry were applied to create a chorologic typology of environmental indicators to map forest extent with an emphasis on large-scale performance, interpretability/communication, and spatial–temporal scalability. Interpretation keys were created to identify forest and non-forest features, and a set of candidate tree cover indices were developed and compared with a decision matrix of prescribed criteria. The candidate indices were intentionally limited to those applying only the visible and NIR bands to obtain the highest possible resolution and be compatible with commonly available multispectral satellites and higher resolution sensors, including aerial and potentially UAV/drone sensors. A new High-Resolution Tree Cover Index (HRTCI) in combination with the Green band was selected as the best index based on scores from the decision matrix. To further improve the performance of the indices, the chorologic typology included two insolation indices, a water index and a NIR surface saturation index, to exclude any remaining spectrally similar but unrelated land cover features such as agriculture, water, and built-up features using a process of elimination. The approach was applied to the four seasons across a wide range of ecosystems in south-eastern Australia, with and without regionalisation, to identify which season produces the most accurate results for each ecoregion and to assess the potential for mitigating the spatial–temporal scaling effects of the Modifiable Spatio-Temporal Unit Problem. Autumn was found to be the most effective season, yielding overall accuracies of 94.19% for the full extent, 95.79% for the temperate zone, and 95.71% for the arid zone. It produced the greatest spatial agreement between two recognised global products, the GEDI forest heights extent and the ESA WorldCover Tree cover class. The performance, transparency, and scalability of the approach should provide the basis for a framework for globally relatable forest monitoring.

1. Introduction

Forest ecosystems are major habitats for terrestrial biodiversity and provide important ecosystem services [1,2]. The consistency and harmonisation of information relating to forest resource variables such as biomass and carbon sequestration stocks are important for global monitoring and biodiversity/conservation efforts, forestry, as well as climate change prediction and mitigation. However, estimates vary substantially. This is largely due to variations in the spatial extent of the different forest masks they have been based on, which are affected by different mapping resolutions and differences in forest definitions [3,4,5,6,7]. There is therefore an increasing need to quantify and reduce these uncertainties [8]. Binary masks reduce the complexity of analyses by separating classes in their simplest semantic forms from other thematically unrelated classes and can facilitate further sub-classification. Forest masks are critical mapping tools for landscape management, biodiversity monitoring and conservation, habitat modelling, timber resource management and wood production, carbon accounting, vegetation mapping, and many other applications.

1.1. Consequences of Different Forest Definitions and Issues with Existing Masks

A major issue affecting forest mapping is the lack of consistency due to the implementation of different forest definitions. Over 800 definitions of forest have been noted worldwide [9], which can be based on vegetation type, composition, altitude, management, and land use objectives, among others [10]. The choice of a forest definition can have a large impact on estimates of extent, deforestation, and forest degradation and on the development of national forest reference emission levels, which serve as benchmarks for climate change mitigation efforts [11].

Some of the technical limitations that exist in the production of forest-related products to date include (1) the need for manual editing [12], a laborious and time-consuming process requiring subjective expertise; (2) limitations in either the spectral [12], spatial, or temporal resolution of remote sensing data [13,14]; (3) the need for inter-annual phenologic observations [12,13,14]; and (4) the need for short wave infrared (SWIR) bands [12,13,14], which are typically at lower resolutions than visible and NIR bands and not commonly available in higher resolution sensors.

Until recently, global forest-related products have typically ranged in pixel dimensions from 100 m to 1 km [15,16]. These low- to moderate-resolution products create mixed classes characterised by a mosaic of trees, shrubs, and herbaceous vegetation, which introduce large variability when they are used to estimate the biomass density in a region and are considered to be too coarse to represent heterogeneous landscapes for some decision-making objectives [17,18]. A comparison of more recent global forest products at higher resolutions, between 25 and 30 m, has shown major disagreements in forested areas in highly fragmented landscapes [19]. Landsat-based studies have a tendency to underestimate forested areas in regions with low forest cover—the effect of sensor spatial resolution on vegetation coverage is particularly pronounced in grassy areas, steppe, and semi-arid woodlands and savannas with sparse, open canopies where forests are prone to be under-represented depending on the forest definition being used, which remains a challenge [12,20]. This has led authors to encourage the use of higher-resolution sensors [12,19].

Global Forest Watch (www.globalforestwatch.org) now uses the term ‘tree cover’ rather than ‘forest cover’ for monitoring global gain or loss to account for non-forest land such as agricultural fruit trees, shrubland, or plantations being incorrectly classified as forest area [21,22,23]. Similarly, the European Space Agency (ESA) WorldCover product defines forests to be ‘Trees observable in the landscape from the images. Parcels planted with fruit trees or shrubs: single or mixed fruit species, fruit trees associated with permanently grassed surfaces. Tree cover percentage classification to >15% and tree height classification to >3 m’ [24]. A spatial comparison of forest extents for Australia, where tree height is cut off at 3 m at a 25 m resolution in the Global Ecosystem Dynamics Investigation (GEDI) forest canopy height dataset for 2019 [14] and the European Space Agency (ESA) WorldCover product Tree cover class (>3 m) is at 10 m for 2020 [25], indicates that the majority of disagreements relative to the ESA product are generally in heterogenous areas, with overestimates using the GEDI product (set at >3 m) in agricultural areas along the coast and underestimates in the more arid and temporally variable landscapes with forest–grassland transitions/ecotones. These problems are typically due to different treatments with mixed pixels, where different land classes occur in the same pixel [26].

Monitoring efforts must address inter-seasonal variation to classify land cover classes correctly across the seasons and avoid false positives in change detection studies. Most landscape indices are sensitive to changing grain and extent with different patterns and processes taking place at certain ‘characteristic scales’ [27]. It is therefore important to properly identify and define characteristic scales within an ecological hierarchy prior to measurement, to ensure the spatial and temporal scale of study is appropriate to the ecological patterns and processes of interest [28,29,30].

The spatial and temporal scaling effects on measurements, inferences, and classifications are known, respectively, as the Modifiable Areal Unit Problem (MAUP) [31,32] and the Modifiable Temporal Unit Problem (MTUP) [33,34]. The MAUP and MTUP have recently been combined to form the ‘Modifiable Spatio-Temporal Unit Problem’ (MSTUP) [35,36], in which the spatial extent, temporal range, and the significance of their clustering vary as their aggregation, segmentation, and boundaries change.

Solutions at a particular scale might not extrapolate accurately to other resolutions [37], so any observation made at a low resolution needs to be tested and confirmed at a higher resolution, and vice versa. An example of the MAUP in remote sensing includes the way in which global vegetation maps often misclassify dry tropical forests as some other class and require higher spatial resolutions to map them correctly and with greater thematic resolution [38]. Regarding the MTUP, consideration must be given to the impacts of both short- and long-term temporal processes including climatic anomalies and natural and anthropogenic disturbances (such as drought, erratic rainfall and flooding, fire, cyclones, logging, and agricultural and urban expansion) which affect average annual reflectance, seasonal variability, and growing season lengths—both during interpretation and accuracy assessments [39].

The MAUP can be most effectively mitigated by reducing analytical complexity with the use of optimal zoning systems or regionalisation [40,41,42]. Regions can be defined at any scale using the way they differ from neighbouring regions in terms of spatial scale, structural attributes, and process rates, and are typically based on spatial changes in the repeating patterns of biotic and abiotic ecosystem properties and inter-functional sets of environmental factors, including climate, geology, landforms, soils, vegetation, and land use [43,44]. Computationally, regionalisation also reduces the data to be processed, as well as the spectral variability and analytical complexity per dataset. Regionalisation should be created at characteristic spatial–temporal scales that recognise both landscape variation and key ecological processes over time [45]. Venkatappa et al. [46] identified phenological periods from time series curves that displayed optimal separations in vegetation index values of land cover classes. In this study, these phenologic ranges are referred to as Moments of Maximum Separability (MoMS) for imagery acquisition to mitigate the effects of the MTUP.

1.2. Aims

Methods that have taken inter-annual phenology into account and applied SWIR bands, including the GEDI Global Forest Canopy Height 2019 dataset [14] and ESA WorldCover, have shown the best results for global forest mapping so far [19]. Downloadable versions of the Global Forest Watch tree canopy cover datasets are only available annually. But how can we produce forest masks with seasonal or monthly image composites for large regions, or single date image acquisitions for local studies? Is it possible to map forest extent consistently across the seasons? If not, then which is the best season to conduct surveys for training and validation data?

High-resolution sensors, such as those of RapidEye, WorldView, and PlanetScope, only have bands that range from the visible (blue, green, red) to NIR bands. For local monitoring, drones can fly under clouds, so they can essentially acquire photography anytime during a year, and currently, it is only common for drones to have sensors with visible and NIR channels, while SWIR channels remain too heavy or costly.

Global mapping should have relevance, utility, consistency, and reliability across scales [21]. Techniques should ideally be as simple and transparent as possible to allow people to easily understand and refine products for their own regions. Scalability is important for harmonising comparable global monitoring systems, and a simple concept and visual interpretability can more easily facilitate the communicability of results. It would be preferable to develop a model for a forest mask with a combination of thematically rationalised and visually interpretable indices that will yield consistent results at different scales and across seasons. A global index requiring only a minimal number of accompanying conditional/masking indices would be ideal to meet these requirements.

The aim of this study was to produce a visually interpretable workflow to create a high-spatial-resolution forest mask that would be scalable across different ecoregions and robust to variations in seasonal conditions. The workflow was intentionally limited to the visible and NIR bands in order to be compatible with higher-resolution sensors including drones, and an effort was made to identify an optimal time of year for single-date image acquisitions (MoMS).

1.3. Hypothesised Model as a Transparent Chorologic Typology of Ecological Gradients

Ecosystems are characterised by interacting and co-varying gradients that vary across space and time due to local to regional biotic interactions, abiotic processes, conditions, and disturbance [47,48]. Gradient analyses represent this variation more consistently with ecological theory than hierarchical patch mosaic models for system organisation because they avoid the need to redefine entities and interactions across spatial and temporal scales, focusing instead on fundamental driving factors [49]. This is an important consideration for the mitigation of the MSTUP and global classification.

In ‘Natural Colour’ RGB images, forests are generally darker the more productive they are because they absorb more radiation and reflect less in the photosynthetically active wavelengths, whereas urban/built-up, agricultural, and desert (or ‘bare land/barren’) features are generally brighter [50,51,52]. This contrast has been referred to as the ‘dark-object concept’ [53]. These properties are represented in shade-related fractional cover ([12,54], etc.) and relative vegetation abundance [55]. Regarding tree cover, trees generally cast shadows, shrubs less so, and grasses rarely at all. To apply the dark-object concept, it is necessary to identify and mask out non-forest dark objects such as water, shadows, dark soils, lava, burn scars, and dark impervious surfaces (such as some urban features), which can be as dark, or even darker, than forest pixels [53]. The generally brighter agricultural features and inland wetlands will have a higher surface albedo than forests. Vegetation in the spectral range of the visible bands (blue to red) displays low reflectance and transmittance due to chlorophyll and carotene absorption, while in the NIR band, it displays high reflectance and transmittance and very low absorption due to internal leaf structures [56]. From visual inspection of the ‘Near-Infrared’ RGB (R,G,B: NIR, red, green), we can also generalise that dark, wet soils, non-forested wetlands, and some urban features are more saturated than forests. Lastly, like forests, water also generally appears dark.

A chorologic typology is a classification of geographic phenomena that represents causal relationships among them [57,58]. Examples of chorologic typologies combining parametric landscape typologies and holistic landscape chorologies have been documented for landscape assessments [59]. Causality can be determined regionally, based on the evidential reasoning from existing landscape ecological knowledge and visual and parametric associations among physiognomic features from available reference data or maps to colourimetric indicators in multispectral false colour images using deductive techniques—such as a process of elimination (for example [60]). These associations can in turn be validated with field-based accuracy assessments.

To develop a simple and transparent tree cover index with high performance, a chorologic typology similar to the Tasseled Cap index [61] is proposed with a set of index-based conditions to exclude other dark objects. A rule-based approach was chosen because of its transparent interpretability [62]. The chorology begins with a tree cover index that is intended to identify forest cover as mutually exclusive as possible and is refined using a process of elimination with a set of selected conditions (or typological descriptors) that are intended to provide the necessary phenologic handling to take into consideration the changing reflectance conditions that will cause different land cover classes to overlap spectrally with forests across the seasons. The hypothesised chorologic typology for the proposed forest mask can be expressed as follows:

Tree Cover Index > a,

And Insolation < b,

And Water false colour Hue < c,

And Surface NIR Saturation > d

The ‘Surface NIR Saturation’ condition is intended to distinguish features such as built-up and agricultural or wetland features that typically display lower saturations in the HSV (hue, saturation, value) colour space [63] compared to forests in RGB composites including the NIR band.

2. Methods

2.1. Study Zone Representativeness

This study was conducted over the extent of the State of New South Wales (NSW), in south-eastern Australia (between 141°2′6.2″E, 37°28′17.8″S and 153°37′7″E, 28°11′25.6″S), which has an area of 801,150 km². NSW is a good testing ground for forest mapping because of its wide variety of vegetation expressions as well as other diverse land cover classes with similar spectral properties over varied environments and terrain. In terms of potential spectrally similar classes, it includes a wide range of vegetation, from rainforests on the coast to shrubs and chenopods in the arid interior; dense urban areas, a wide range of agricultural features, including intensely irrigated agricultural areas; a range of freshwater and saline wetlands on the coast and inland; and an alpine range with snow cover during the colder months in the south. In terms of biome representativeness, it covers the area in the red polygon in the Whittaker Biome ordination [64], as shown in Figure 1, which more specifically, according to Dinerstein et al. [65], includes temperate broadleaf, mixed forests, grasslands, savannas and shrublands, Mediterranean forests, woodlands and scrub, and desert and xeric shrublands.

2.2. Phenological Knowledge Base

Australian climates display high intra- and inter-annual variability, largely due to the combined circulation patterns of the El Niño Southern Oscillation (ENSO) and the Indian Ocean Dipole [45,66,67,68]. Australian ecosystems, therefore, exhibit marked differences in seasonal patterns in canopy growth and plant life cycle events compared with the more predictable seasonal phenology of temperate deciduous and boreal regions. Some coastal temperate vegetation may include greening and senescing events in their temperate eucalypt understorey but display constant moderate-to-high productivity with less temporal variability than the drier inland [69]. West of the Great Dividing Range (mountain range), the southern agricultural areas display both seasonal and non-seasonal variation, the northern agricultural areas display relatively high greenness and strong non-seasonal variation, and the riparian areas further west display relatively strong but non-seasonal variations in productivity [69]. Further west from the semi-arid to arid west, the dominant vegetation types are grasses and shrubs, which display high inter-annual phenologic variability depending on rainfall [70]. Summer has typically been recognised to provide better spectral separability of forests in Australia, when grasses are generally senescent, providing better contrast with the greener overstorey foliage [12,71,72]. This also applies to savannas around the world, where it is more appropriate to consider the ‘dry season’ or ‘end of season’ [73,74].

2.3. Eco-Geographic References

The state of NSW has a wealth of existing reference data. The main references for this study were a botanical framework [75], and the corresponding vegetation map [76], which were composed as a mosaic of existing regional maps quite generally at a resolution of 200 m. Additional reference maps included the Landuse Mapping for NSW 2017, v1.2 [77] for agricultural features, the mapping from Seamap Australia for detailed coastal wetland features (particularly the Estuarine macrophytes dataset) [78], and the ESA’s WorldCover product for confirmation of general land cover classes.

2.4. Analysis Design

The holistic interpretation of a class within part of an image depends on the interpretation of the rest of the image and surrounding landscapes [79]. Holistic reduction addresses the need to balance holism and reductionism in ecological studies in order to explain outcomes by looking at parts of complex systems against the desire to understand how the parts work together in a fully functioning system [80]. A holistic reductionist approach to remote sensing [60,81] was undertaken to reduce analytical complexity by interpreting a single reduced class with an emphasis on large-scale performance, scalability, and global transferability.

A pixel-based analysis was applied for the greatest precision, transparency, and processing speed, which is not possible when processing intensive segment-based multivariate correlation approaches. This is particularly so for large-scale mapping efforts that encompass semi-arid to arid areas and savannas with sparse and varied tree cover. Tree cover was defined to represent trees greater than or equal to 3 m in height, including tall shrubs, forest plantations, and fruit trees in order to harmonise with the globally recognised ESA WorldCover and GEDI datasets, both of which are publicly available on Google Earth Engine. The ESA product was produced with the 3 m height definition, whereas the GEDI height product (unlike the Hansen Global Forest Change products [21]) can have its height set. To improve spatial and temporal resolution, Sentinel-2 imagery was used for its 10 m resolution visible and NIR bands with a revisit time of 5 days. This study was conducted in the phases described below.

2.4.1. Phase 1: Selection of an Indicative Inter-Annual Range and Imagery Preparation

Inter-annual, median-based image composites were generated using Google Earth Engine (GEE) for each season spanning the years 2016 to 2018, to simulate the typical seasonal variability expected during an ENSO event and mitigate the influence of the ‘Black Summer’ bushfires that occurred between 2019 and 2020. The full Sentinel-2 Top-Of-Atmosphere reflectance Level 1C archive was considered, after removing scenes with a cloudy pixel percentage of more than 20%. Dense and cirrus clouds were also masked out using a per-pixel approach, with the bitmasks provided in the Level 1C processing algorithm. This process provided a seamless, well-colour-balanced imagery mosaic indicative of the typical conditions expected during those seasons.

2.4.2. Phase 2: Regionalisation by Ecoregions

Rainforests and wet and dry sclerophyll forests are predominantly located in the east of New South Wales (Figure 2), while semi-arid woodlands and arid shrublands are predominantly located in the west [75]. An aggregation of ecoregions [65] adequately defines the boundaries for these major zones (Figure 2). The Australian Alps montane grasslands ecoregion was omitted from this study in order to concentrate on tree cover by excluding variations due to snow cover.

2.4.3. Phase 3: Evaluation of Alternative Tree Cover Indices

Interpretation Keys with Colourimetric Benchmarks

Interpretation keys with colourimetric benchmarks were developed for the variety of known forest types (Figure 3) and a range of non-forest features with similar spectral properties to forests (Figure 4) around the study zone as a basis for interpretation. The benchmarks were identified visually and selected for the variety of hues and saturations they exhibited. The interpretations were drawn from the map references (see the eco-geographic references section) and are described in the keys. Deductions were assisted with Google Street View as virtual field visits where they were available to confirm the presence of trees, and the vectorized information from Google Maps was used to confirm the presence and extent of semantically ambiguous features such as fruit trees in contextually identifiable farms.

The general pattern that can be observed from the SII key of colourimetric benchmarks in the ‘Land/Water’ RGB for forests is that rainforests on the coast appear orange, sclerophyll forests and forestry plantations in the eastern zone coast range from reddish and maroon tones to greyish teal tones, while fruit trees appear bright orange or red. A colourimetric gradient from red to greenish-cyan was deduced from a holistic scan of the study zone to represent vegetation productivity in the ‘Land/Water’ RGB in GEE.

Most of the non-forest features are obviously different colours compared to those from the forest SII key; however, an interpretation of texture, saturation, and spatial context or visualisation in natural colour is necessary to distinguish some features with greenish tones.

Development and Comparison of Candidate Indices with a Decision Matrix

Taking into consideration the dark-object concept, the first two multispectral bands, i.e., blue and green, appear to show the most contrast between forests and the brighter features; however, they also display the highest radiometric imbalances across satellite acquisition flight paths. The red band appears more balanced, but it noticeably displays agricultural features in the same ranges as forests in a pseudo-colour stretch. A simple band ratio or a subtraction between the green and blue bands reduced the radiometric imbalances and appeared to separate most agricultural features and urban buildings from woody vegetation, but it omitted drier rainforests and also misclassified water and roads as tree cover. They also removed the expected productivity gradients across the landscape. To ameliorate this, a combination of the NIR band with the band ratio subtraction: NIR/green minus NIR/blue was tested. This index filled in the missing drier rainforests and displayed a recognisable productivity gradient. Progressive modifications to the band ratio subtraction were tested, as listed in

Table 1. Shortlist of candidate tree cover indices. Where B1 = blue, B2 = green, B3 = red, B4 = NIR.

Equation	Number
NDVI = (B4 − B3)/(B4 + B3)	(1)
B2	(2)
(B1 − B2)/B3	(3)
(B4/B2) − (B4/B1)	(4)
(B4/B2 − B4/B1)/B2	(5)
(B4/B2 − B4/B1)/(B3 + B2)	(6)
(B4/B2 − B4/B1)/(B3/B2)	(7)
(B4/B2 − B4/B1)/(B3 + B2 + B1)	(8)
((B4/B2 − B4/B1)/B3) − ((B4/B2 − B4/B1)/B4)	(9)

. The testing was limited to non-parametric algebraic indices in order to (1) promote scalability since statistically derived indices with fixed coefficient weightings from a limited number of samples for a particular area are unlikely to perform well in other areas and (2) to facilitate backward and existing cross-sensor compatibility and calibration.

The development of the candidate indices was guided by a holistic visual assessment of the omission or commission of the colourimetric benchmarks specified in the forest and non-forest SII keys with a decision matrix as a process for refinement [60,81]. For every seasonal image composite derived from Sentinel-2 data, each index was individually analysed and evaluated in a GIS at 100 m resolution. The evaluation involved density slicing of quantiles in pseudo-colour, resulting in the identification of a shortlist of the best-performing indices. The shortlisted indices were then further validated at full resolution using GEE. The manual classification allowed for visual deductions to be verified with aerial photography interpretation (API). This involved overlaying the resulting classifications on the very-high-resolution (VHR) imagery in GEE when the presence of tree cover was not obvious, or when there were obvious errors in the reference maps. The tree cover indices with the highest scores in the decision matrix were used for further testing. These were the green band by itself and index number 6 (Table 1) that will herein be referred to as the HRTCI (High-Resolution Tree Cover Index):

HRTCI = (NIR/Green − NIR/Blue)/(Red + Green)

Determining Index Thresholds with Pseudo-Invariant Features (PIFs)

Pseudo-invariant features (PIFs) are regularly used to create pseudo-training and validation datasets to reduce the cost and difficulty of collecting field-based ground truth training data [82]. Geographers frequently distinguish between the core and periphery of features, and most image analysts agree on the locations and characteristics of the cores of features [42], whereas the periphery of features is more uncertain and typically produces mixed pixels due to ecotones and edge effects. Using gradient-directed indices, the primary concern for forest masking was to identify the extremes of forests at the transition into treeless areas for each zone. In semi-arid and arid ecoregions, woodland trees and tall shrubs may occur in a variety of landscape contexts, including slightly more elevated areas or uplands (slopes, dunes, and hills), in frequently flooded sites including drainage channels, around the margins of ephemeral lakes and intermittent streams and creeks, and scattered across drier depressions and flats throughout riverine and flood plains [75]. To narrow down the likely spatial range for these extremes, the 3 m interval from the GEDI forest heights dataset was used to identify areas likely to have trees with the shortest heights, and the aforementioned botanical knowledge of the location of forest extremes was used to define three visually interpreted PIFs for transitions into treeless vegetation in the western zone and one PIF in the eastern zone (Figure 5). Tree cover presence in these areas was confirmed with API if the features that conformed to the SII forest key appeared to cast shadows. The pseudo-colour quantile intervals for the candidate indices were therefore classified as tree cover until they reached the designated PIFs.

2.4.4. Phase 4: Development of a Colourimetrically Assisted Chorologic Typology

Existing vegetation indices such as the NDVI are unable to create accurate forest masks on their own, even during dry seasons or the summer, when grasses and agriculture are thought to produce the most contrasting reflectance compared to woody vegetation [12,71,72]. While the NDVI may show a generalised gradient of productivity across varying spatial and temporal scales [83], it will group forests and agricultural features with similar photosynthetic capacities in the same ranges along the index [84]. It is therefore unsuitable for inferring vegetation fraction at different resolutions over heterogeneous surfaces because of its nonlinearity and scale effects over partially vegetated surfaces with darker soils and shadows [85].

Forest masking requires at least a pair of indices to approximate separability, and the selection of accompanying indices will vary depending on ecoregions [86]. Colourimetrically assisted classification has been shown to facilitate interpretation and communicability and significantly reduce (or even eliminate) sampling effort [60,81]. Colourimetric indices have therefore been proposed for three environmental indicators: (1) insolation, (2) water extent and (3) surface NIR saturation.

Two colourimetric insolation indices were applied. Insolation index 1: The natural colour RGB’s intensity in the HSI (hue, saturation, intensity) colour space:

Blue + Green + Red

Insolation index 2: A variation of the insolation index 1, including the NIR band, which is commonly referred to as brightness in the object-based image analysis sector, expressed as:

Blue + Green + Red + NIR

A previous study on water extent mapping determined that the hue and a modified saturation from an HSV transformation of a normalised difference RGB provided an effective solution to high-resolution water mapping for sensors limited to visible and NIR bands [81]. It was shown to be capable of classifying the full range of water types from clear to turbid in a single index without misclassifying vegetation as water:

NDCHRWI (Normalized Difference Colourimetric High Resolution Water Index) =

R,G,B: ((Red − NIR)/(Red + NIR)) + 1,

((Blue − Green)/(Blue + Green)) + 1,

((Green − NIR)/(Green + NIR)) + 1

→ H,S,V: Hue

A modified saturation index for the NDCHRWI RGB was intended to mask out highly saturated features such as bright urban features, dark and wet agricultural soils, and snow. It was created using a ratio between the minimum and maximum values of each pixel:

NDCHRWI RGB modified saturation = Min(R,G,B)/Max(R,G,B)

Two indices were selected from the decision matrix: the HRTCI and the Green band by itself. The full forest masking chorologic typology can therefore be expressed as:

[(NIR/Green − NIR/Blue)/(Red + Green) > a, OR Green < b],

AND [(Blue + Green + Red) < c, OR (Blue + Green + Red + NIR) < d],

AND NDCHRWI → H,S,V: H < e

AND NDCHRWI Min(R,G,B)/Max(R,G,B) < f

where the letters a to f are index threshold values.

Figure 6 provides a visualisation of a manual classification for the chorologic typology indicators on a summer seasonal image composite. It shows the distribution of each indicator in pseudo-colour and highlights the trade-off in results that would be made by classifying the whole extent compared to classifying the eastern zone separately.

There are noticeable improvements for the eastern zone when it is classified separately for the HRTCI, insolation 2, and the NDCHRWI-modified saturation. Visually relating the ESA WorldCover land classes reference map to the reference imagery, built-up, grasslands, cropland and bare/sparse vegetation appear brighter than tree cover in the ‘Land/Water’ RGB. The thresholds for the HRTCI, the Green band and the insolation indices were determined predominantly with the visual cue of brightness and where they passed the density slicing intervals where the PIFs were defined. The water bodies were interpreted contextually and by hue. They were generalised to range from black to dark blue, to blue to purple and magenta in the ‘Land/Water’ RGB. And the high saturation features for the NDCHRWI-modified saturation index can be visually recognised as appearing either blue or dark blue in the ‘Landscape’ RGB.

2.4.5. Phase 5: Spatial–Temporal Comparison of the Proposed Chorologic Typology

To compare the results with or without regionalisation and to indicate the spatial–temporal uncertainty in the proposed chorologic typology, comparisons were conducted for random forest classifications for each seasonal composite and an annual one in (1) the full extent of the study area, (2) the eastern zone, and (3) the western zone.

For the sampling stratification, 10,000 points were assigned across the full extent of the study area, with 50% for each of the two zones. For each spatial extent, 50% of the points were allocated to forest and 50% to non-forest areas as agreed upon using an intersection of the GEDI tree height dataset and ESA WorldCover ‘Tree cover’ class. The forest points were distributed with a randomised areal stratification of 5 quantiles from the GEDI dataset heights, while the non-forest points were distributed with a randomised areal stratification of the ESA WorldCover classes. A visual appraisal of both these reference datasets determined that neither was accurate enough to be used as a reliable reference for tree cover extent in urban areas. These areas were therefore masked out using the World Settlement Footprint 2015 [87]. The points were combined for the classification of the full extent; 50% of the combined points were used for training in GEE, while 50% were used for validation with the accuracy assessment in a GIS.

3. Results

3.1. Evaluation and Selection of Tree Cover Indices with a Decision Matrix

All of the newly proposed indices demonstrated fewer errors of commission compared to the NDVI. Only some of the indices displayed a consistent spatial distribution of quantiles relative to the expected distribution of vegetation formations across the seasons, suggesting that some indices can be more stable than others. For particular seasons, some indices displayed errors of commission, such as agricultural features or the highly irrigated soils and errors of omission for hill-shaded forested areas. It was therefore necessary to assess all seasons to identify all potential symptomatic issues.

Index 9 provided the greatest differentiation of vegetation formations during the summer (

Table 2. Candidate tree cover index decision matrix. Colour-related descriptions refer to visualisation in the ‘Land/Water’ RGB: NIR, SWIR1, red.

	Criteria	1	2	3	4	5	6	7	8	9
Overlap of spectrally similar features	Overlaps clear water (−1)	0	−1	−1	−1	−1	−1	−1	−1	0
	Overlaps turbid water (−1)	0	0	0	0	0	0	0	0	−1
	Overlaps built-up features (buildings, roads & mining) (−1)	−1	0	−1	−1	−1	−1	−1	−1	−1
	Overlaps orange agriculture with sclerophyll range in Summer or Winter (−1 if moderately, −2 if excessively)	−2	0	0	0	0	0	0	0	0
	Overlaps blue agriculture excessively in one or more seasons (−1)	0	0	−1	0	0	0	−1	0	0
	Overlaps bright greenish agriculture excessively in one or more seasons (−1)	−1	0	0	0	0	0	0	0	0
	Overlaps coastal wetlands (particularly sedgelands) (−1)	−1	−1	−1	−1	−1	−1	−1	−1	−1
	Overlaps inland wetlands (−1)	0	0	−1	−1	−1	−1	−1	−1	−1
Gradient and Seasonality	Gradient conforms to local reference of vegetation formations across seasons (+1 if generally or for most seasons, +2 if for all seasons)	1	1	1	1	1	2	1	2	2
	Displays variation in quantile distribution across the seasons (−1)	0	0	−1	0	0	0	−1	0	0
Illumination effects	Displays erroneous hill shading effects in mountainous areas during Winter, with forested hill shades classified as non-forest (−1 if minimally, −2 if excessively)	−2	0	0	−2	0	0	0	−1	−1
	Displays atmospheric illumination imbalances across satellite flight paths (−1)	0	−1	0	0	−1	0	0	0	0
	Score:	−6	−2	−5	−5	−4	−2	−5	−3	−3

), with more intervals for forest types in the eastern zone, which is characterised by rainforest and wet and dry sclerophyll forests. However, it misclassified large turbid water bodies in different seasons in the arid west, and as with index 8, displayed albeit small amounts of omissions in heavily hill-shaded areas. Indices 2 and 6 scored the highest (Table 2) but for different reasons. Index 2, the Green band by itself, clearly displayed the fewest errors of commission across all the seasons, but it displayed high BRDF (bidirectional reflectance distribution function) [88] effects across satellite image flight paths and did not provide a detailed gradient. As a consequence, its results did not conform to the known distribution of regional vegetation formations. Conversely, index 6 (HRTCI), provided a strong and relatively consistent gradient distribution across all seasons with errors of commission that could be removed with masking from the Green band or the colourimetric indices in the proposed chorologic typology.

3.2. Seasonal Accuracy Comparison of the Proposed Chorologic Typology

All accuracy estimates were relatively high (>92%,

Table 3. Summary of accuracy assessments for the chorologic typology for each season at different extents, run with random forest classifications.

		User Accuracy (%)		Producer Accuracy (%)		Overall Accuracy (%)	Ranking
	Season	Tree	Non-Tree	Tree	Non-Tree
Full extent	Summer	93.14	93.63	93.52	93.26	93.39	4
	Autumn	93.82	94.55	94.46	93.92	94.19	2
	Winter	93.89	93.67	93.49	94.05	93.78	3
	Spring	93.24	92.73	92.51	93.44	92.98	5
	Annual	94.45	94.29	94.14	94.59	94.37	1
Eastern zone	Summer	95.40	94.79	94.83	95.36	95.09	5
	Autumn	96.19	95.40	95.43	96.17	95.79	2
	Winter	95.18	95.66	95.75	95.08	95.41	3
	Spring	95.67	94.88	94.91	95.64	95.27	4
	Annual	95.92	96.19	96.26	95.84	96.06	1
Western zone	Summer	93.47	94.34	94.38	93.43	93.90	3
	Autumn	95.52	95.90	95.90	95.51	95.71	1
	Winter	93.99	93.14	93.05	94.07	93.56	4
	Spring	92.49	92.58	92.57	92.51	92.54	5
	Annual	94.81	94.76	94.74	94.83	94.78	2

). Autumn consistently provided the highest accuracy for each of the extents (Table 3); however, the annual composites provided slightly better results for the full extent and the eastern zone by itself. Autumn can thus be considered the MoMS for optimum imagery acquisition. Accuracy estimates for the best-performing seasons and the annual composites in the individual zones were higher than those in the full extent (Table 3). All seasons and the annual composite benefited from the regionalisation for the eastern zone; however, this only applied to the summer, autumn, and the annual composite for the western zone.

The detailed accuracy assessments stratified by ESA WorldCover land cover classes (

Table A1. Accuracy assessment for the full extent for the summer season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	321	100	3523	978	54	36	29	1	5042	93.63
	Tree cover	4629	29	272	25	2	1	8	4	4970	93.14
	Total	4950	129	3795	1003	56	37	37	5	10,012
	Producer accuracy	93.52	77.52	92.83	97.51	96.43	97.30	78.38	20
		93.52	93.26
Overall accuracy:											93.39

,

Table A2. Accuracy assessment for the full extent for the autumn season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	274	106	3547	983	53	36	28	1	5028	94.55
	Tree cover	4676	23	248	20	3	1	9	4	4984	93.82
	Total	4950	129	3795	1003	56	37	37	5	10,012
	Producer accuracy	94.46	82.17	93.47	98.01	94.64	97.30	75.68	20
		94.46	93.92
Overall accuracy:											94.19

,

Table A3. Accuracy assessment for the full extent for the winter season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	322	106	3564	970	55	36	30	0	5083	93.67
	Tree cover	4628	23	231	33	1	1	7	5	4929	93.89
	Total	4950	129	3795	1003	56	37	37	5	10,012
	Producer accuracy	93.49	82.17	93.91	96.71	98.21	97.30	81.08	0
		93.49	94.05
Overall accuracy:											93.78

,

Table A4. Accuracy assessment for the full extent for the spring season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	371	108	3545	958	55	37	27	0	5101	92.73
	Tree cover	4579	21	250	45	1	0	10	5	4911	93.24
	Total	4950	129	3795	1003	56	37	37	5	10,012
	Producer accuracy	92.51	83.72	93.41	95.51	98.21	100	72.97	0
		92.51	93.44
Overall accuracy:											92.98

,

Table A5. Accuracy assessment for the full extent for the annual image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy	(Could not be processed)
Classified data	Non-tree cover	290	107	3589	976	54	36	26	0	5078	94.29
	Tree cover	4660	22	206	27	2	1	11	5	4934	94.45
	Total	4950	129	3795	1003	56	37	37	5	10,012
	Producer accuracy	94.14	82.95	94.57	97.31	96.43	97.30	70.27	0
		94.14	94.59
Overall accuracy:											94.37

,

Table A6. Accuracy assessment for the eastern zone for the summer season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	130	3	1619	684	23	27	5	2	2493	94.79
	Tree cover	2386	3	94	10	2	0	3	3	2501	95.40
	Total	2516	6	1713	694	25	27	8	5	4994
	Producer accuracy	94.83	50	94.51	98.56	92	100	62.50	40
		94.83	95.36
Overall accuracy:											95.09

,

Table A7. Accuracy assessment for the eastern zone for the autumn season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	115	6	1636	687	24	27	2	1	2498	95.40
	Tree cover	2401	0	77	7	1	0	6	4	2496	96.19
	Total	2516	6	1713	694	25	27	8	5	4994
	Producer accuracy	95.43	100	95.50	98.99	96	100	25	20
		95.43	96.17
Overall accuracy:											95.79

,

Table A8. Accuracy assessment for the eastern zone for the winter season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	107	6	1614	680	24	26	5	1	2463	95.66
	Tree cover	2409	0	99	14	1	1	3	4	2531	95.18
	Total	2516	6	1713	694	25	27	8	5	4994
	Producer accuracy	95.75	100	94.22	97.98	96	96.30	62.50	20
		95.75	95.08
Overall accuracy:											95.41

,

Table A9. Accuracy assessment for the eastern zone for the spring season image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	128	5	1632	680	24	27	2	0	2498	94.88
	Tree cover	2388	1	81	14	1	0	6	5	2496	95.67
	Total	2516	6	1713	694	25	27	8	5	4994
	Producer accuracy	94.91	83.33	95.27	97.98	96	100	25	0
		94.91	95.64
Overall accuracy:											95.27

,

Table A10. Accuracy assessment for the eastern zone for the annual image composite.

		Reference Data										Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Mangroves	Total	User accuracy
Classified data	Non-tree cover	94	6	1627	689	24	27	2	0	2469	96.19
	Tree cover	2422	0	86	5	1	0	6	5	2525	95.92
	Total	2516	6	1713	694	25	27	8	5	4994
	Producer accuracy	96.26	100	94.98	99.28	96	100	25	0
		96.26	95.84
Overall accuracy:											96.06

,

Table A11. Accuracy assessment for the western zone for the summer season image composite.

		Reference Data									Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Total	User accuracy
Classified data	Non-tree cover	140	113	1820	325	51	3	20	2472	94.34
	Tree cover	2349	21	119	22	1	1	0	2513	93.47
	Total	2489	134	1939	347	52	4	20	4985
	Producer accuracy	94.38	84.33	93.86	93.66	98.08	75	100
		94.38	93.43
Overall accuracy:										93.90

,

Table A12. Accuracy assessment for the western zone for the autumn season image composite.

		Reference Data									Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Total	User accuracy
Classified data	Non-tree cover	102	118	1855	336	51	4	20	2486	95.90
	Tree cover	2387	16	84	11	1	0	0	2499	95.52
	Total	2489	134	1939	347	52	4	20	4985
	Producer accuracy	95.90	88.06	95.67	96.83	98.08	100	100
		95.90	95.51
Overall accuracy:										95.71

,

Table A13. Accuracy assessment for the western zone for the winter season image composite.

		Reference Data									Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Total	User accuracy
Classified data	Non-tree cover	173	115	1830	329	51	4	19	2521	93.14
	Tree cover	2316	19	109	18	1	0	1	2464	93.99
	Total	2489	134	1939	347	52	4	20	4985
	Producer accuracy	93.05	85.82	94.38	94.81	98.08	100	95
		93.05	94.07
Overall accuracy:										93.56

,

Table A14. Accuracy assessment for the western zone for the spring season image composite.

		Reference Data									Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Total	User accuracy
Classified data	Non-tree cover	185	117	1801	316	51	4	20	2494	92.58
	Tree cover	2304	17	138	31	1	0	0	2491	92.49
	Total	2489	134	1939	347	52	4	20	4985
	Producer accuracy	92.57	87.31	92.88	91.07	98.08	100	100
		92.57	92.51
Overall accuracy:										92.54

and

Table A15. Accuracy assessment for the western zone for the annual image composite.

		Reference Data									Index Importance Ranking
	Classes	Tree cover	Shrub land	Grass land	Crop land	Bare/sparse vegetation	Water	Herbaceous wetlands	Total	User accuracy
Classified data	Non-tree cover	131	115	1852	325	51	4	20	2498	94.76
	Tree cover	2358	19	87	22	1	0	0	2487	94.81
	Total	2489	134	1939	347	52	4	20	4985
	Producer accuracy	94.74	85.82	95.51	93.66	98.08	100	100
		94.74	94.83
Overall accuracy:										94.78

in Appendix A) indicated that shrubland displayed proportionally more errors than grassland, cropland, and bare/sparse vegetation for the full extent and the western zone. The extent of shrubland in the eastern zone was too scarce to make any inferences. The same can be said for the representativeness of herbaceous wetlands and mangroves along the coast, where there was a very limited areal agreement between the GEDI and ESA reference datasets. While also relatively limited in sample size, herbaceous wetlands in the western zone scored well.

In the eastern zone, the HRTCI ranked highest in importance for each of the seasons except for winter. However, in the western zone, NDCHRWI hue ranked highest for the summer and autumn, while the first variation of the insolation index ranked highest in the winter and spring. For the full extent, NDCHRWI hue ranked highest for all seasons except for autumn, the highest scoring season, where HRTCI ranked highest.

3.3. Comparison of Proposed Forest Masking with GEDI and ESA Using the Best-Performing Seasonal Composites

Spatial comparisons of agreement and disagreement with the reference layers specified in the eco-geographic references section indicate that the new chorologic typology during the best performing season (autumn) differentiate urban features better than the GEDI and ESA products (Figure 7 and Figure 8). While there are disagreements between GEDI and ESA for different agricultural areas, the chorologic typology clearly displays misclassification errors in highly irrigated cropland areas (appearing blue in the overlays for both GEDI and ESA and rationalised to be irrigated croplands from the Landuse Mapping for NSW 2017, v1.2). The chorologic typology, like GEDI, also displays misclassifications in hill-shaded areas in both zones in the autumn composite, particularly for grasslands and pastures in the eastern zone. However, the problem was reduced between late spring and early summer when hill shading is less pronounced. The chorologic typology is able to map fine features, such as isolated trees and small forest patches, riparian forest strips, and tree fences across both zones, including more individual isolated trees than both GEDI and ESA (Figure 7, examples 2 and 3, and Figure 8, example 2). The chorologic typology clearly misclassifies coastal wetland features, particularly sedges. However, it appears to map arid wetland features more comprehensively than both the GEDI and ESA products.

4. Discussion

The proposed chorologic typology has built upon previous studies that presented applications of colourimetry [60] and holistic reduction [81] to remote sensing. The results were modestly improved with an ecoregional stratification, which reduced the multispectral variation in land cover features across the full extent of the study area. Regionalisation (or areal differentiation) mitigated the effects of the MAUP using optimizing index thresholding because the distribution of indices is a relative measure determined by the composition, complexity, and inter-relationships among phenomena across different landscapes or ecoregions [57,58].

The results did not support the common perception that summer is the best season for image acquisition. Although this may be true for individual indices, it did not hold true for a selection of indices in combination. Applying the chorologic typology, the summer was in fact the lowest-scoring season in the eastern zone. For south-eastern Australia, autumn was the MoMS that best mitigated the MTUP. The diversity of forest representation across the study area demonstrated the chorologic typology’s consistency and scalability across different ecoregions.

The HRTCI generally ranked highest in the moister, predominantly temperate regions in the eastern zone, while the NDCHRWI hue ranked highest in the more arid western zone. This result could be related to the difference in temperature and rainfall limitations on vegetation between the respective ecoregions.

The workflow can readily be applied anywhere in the world. Regional SII keys can allow people to communicate and replicate the interpretive process to validate and refine the indicators for their areas [60,81]. Comparison of outputs from SII using an overlay with generally accepted global datasets such as the GEDI and ESA products (that are known to have errors of their own) will allow users to scrutinise the maps in detail with API to evaluate results or establish a diagnostic consensus with other analysts about any systematic failings. A chorologic typology facilitates the conceptual understanding and visualisation for a process of elimination; however, results from a chorologic typology are expected to be more accurate with an automated classification such as a random forest classifier rather than a manual one. Interpretability is one of the most desirable qualities of rulesets, which provides end-users with confidence in the decisions made using a model. While random forests can achieve higher accuracy, they do so at the expense of the intrinsic interpretability present in rulesets [89]. Simple chorologic typologies created from a limited number of high-precision indices are recommended in order to promote simplicity, efficiency, and communicability. They can in turn provide inputs of class extents for further sub-classification using any preferred technique.

An index based only on the bands available at higher spatial resolution (blue, green, red, and NIR) and mapping at a higher resolution of 10 m (compared with the 25 m resolution mapping from the GEDI product), improved tree cover mapping in more arid areas where tree cover is more sparse and tree crowns are smaller, and made it possible to detect and map isolated trees and small forest patches, riparian forest strips, and remnant and planted tree corridors. These fine features can have some ecological and conservation values [23], for example, connectivity, so their inclusion would benefit some applications, while other applications (e.g., reporting on vegetation conversion rates) require them to be excluded.

Alvarez-Vanhard et al. [90] identified the potential ecological insights that multiscale explanations can provide with data fusion and interoperability between very-high-spatial resolution imagery from drones/UAVs and large-scale time series data from satellite-based sensors. While the visible and NIR bands from multispectral sensors are not directly comparable with those of drones/UAVs [91], it is expected that the results from the equivalent Sentinel-2 bands from this study could translate to drones/UAVs in the future once the necessary radiometric inter-calibration, upscaling, testing, and refinement based on solar radiation conditions are conducted [92,93,94]. The indices developed here are compatible with the new generation of high-resolution satellites such as those from PlanetScope, which can provide daily acquisition for dense time series change detection studies at 3 m resolution [95]. They would also allow for retrospective time series analyses to be conducted with archival imagery from the SPOT, RapidEye, and WorldView satellites and aerial photography from ADS40, for example.

The quality of available reference datasets will determine the quality of the results from any analysis. Any errors in the reference data will be inherited by the selected model. Errors may be due to differing land cover class definitions, low spatial resolution and the classification accuracy of reference data, differing time of data collection and classification, and the prevalence of spectrally similar classes [82]. Sample datasets that include mixed pixels can also decrease the accuracy of algorithms and provide erroneous validations [96].

The areal agreement of coastal wetland areas between the GEDI and ESA datasets was very limited in proportion to that of the other land cover classes. An additional visual appraisal suggested that the chorologic typology may not have sufficed to well separate forests from wetland sedges shorter than 3 m in all seasons. Further investigation is required to validate the performance with more points from local datasets to decide whether it would be necessary to develop an additional masking index to separate these particular wetland features, which may require SWIR bands.

Further quality assurance is recommended for the forest mask produced here. The comparisons with trusted reference datasets have been very useful, but the reference datasets have their own errors as well as resolution issues that limit comparisons. The GEDI product was published at 25 m resolution, whereas the ESA data was at 10 m, and it is important to recognise the geo-positional displacements between VHR reference imagery/photography, the Sentinel-2 source imagery, and the reference maps (GEDI and ESA) to explain some expected areal discrepancies. Field sampling for subsequent mapping refinements for the chorologic typology should prioritise the areas of disagreement for forest and non-forest features between the existing reference datasets and the areas of disagreement for the highest-performing seasons from preliminary classification efforts as well as additional extreme PIFs for extent mapping. Further testing with these additional data could also resolve whether or not results would be improved by classifying other land cover classes more explicitly (as with the ESA land cover product) compared with a binary classification (as with the GEDI product and the chorologic typology presented here). Further testing should investigate whether additional sub-regionalisation would improve the results. Additional purposive sampling in areas displaying high levels of hill shading between autumn and spring should also help to reduce misclassifications without the need for DEM indices, which are still not feasible for drone mapping.

The proposed workflow has the potential to be upscaled for a global mapping effort, which would require a global regionalisation rationalised using characteristic scales determined by ecoregional biophysical constraints. The chorologic indices presented in this study were intended as inputs for the validation of knowledge-based perceptual deductions with a random forest classification; however, they could be processed using any statistical modelling or machine learning procedures. The indices are expected to be globally indicative; however, they may need to be weighted to account for the specific environmental constraints of other ecoregions, or additional indices could be applied to exception handle any particularities that may not have been present in the study area.

Compared with indices based on ratios, indices that are based on additions, such as the two insolation indices applied in this study, are more prone to compounding the radiometric imbalances that cause differences in brightness across satellite flight paths, known as the BRDF (bidirectional reflectance distribution function), particularly in humid tropical areas beyond this study area. It is therefore recommended that an orbital normalisation should be processed on the median-based image composites in these areas to guarantee a smooth BRDF correction [88].

To apply the proposed chorologic typology to monitoring forests in alpine environments, it would also be necessary to track the presence and absence of snow accurately. A High-Resolution Snow Index (HRSI), also limited to visible and NIR bands, was developed as part of concurrent research to classify snow distinctly without the misclassification of water and forested hill shadows.

HRSI = Green/((Red + Blue)/2)/Blue

Coarse seasonal timeframes were used to infer MoMS in this study. However, it may be that particular ranges such as from mid-summer to mid-autumn, for example, or even a single month (if a cloudless composite can be obtained) might produce stronger separations. These temporal ranges will also vary latitudinally. The selection of MoMS ranges might be automated with land surface phenology methods, such as maximum separation [97]. This applies a moving window to estimate the ratio of observations that exceed a given threshold to establish the start and end of vegetation growing seasons, which can be applied directly to high temporal density, non-smoothed time series data.

5. Conclusions

Higher spectral and temporal resolutions and increased accessibility to imagery at larger scales have permitted studies that take phenology into account to produce higher accuracies with greater thematic resolutions across scales by differentiating a greater variety of landscape changes [39,43]. The new forest masking workflow produced in this study demonstrates the utility of a simple, conceptually explainable chorologic typology to address the wide variation in tree form, sparsity, and phenology across a large area of interest spanning multiple ecoregions. The results indicated that it is possible to map the fine details of tree cover with high-resolution sensors limited to visible and NIR bands for single-date or high-temporal imagery/photography acquisitions.

Regionalised MoMS-based, single-image composites can be identified in order to promote contrast, preserve detail, and reduce the spectral complexity of datasets due to phenologic and physiognomic variations in order to reduce the effects of the MSTUP (Modifiable Spatio-Temporal Unit Problem). These images can be considered minimum investment purposive satellite imagery sampling designs that can significantly simplify analyses and maximise accuracy for large scale, high resolution, and regionally relatable land cover mapping. Implementation and scaling considerations suggest that the model could be extended to a continental or even a global scale for ecoregionally comparative tree cover mapping and monitoring.