CLASSIFYING DOMINANT PARKLAND SPECIES IN A WEST AFRICAN AGROFORESTRY LANDSCAPE USING PLEIADES SATELLITE IMAGERY

(1)

0

Department of Thematic Studies Environmental Change

MSc Thesis (30 ECTS credits) Science for Sustainable development

Simon Lunn

Reviewer Comments Version 11/05/2020

CLASSIFYING DOMINANT PARKLAND

SPECIES IN A WEST AFRICAN

AGROFORESTRY LANDSCAPE USING

PLEIADES SATELLITE IMAGERY

(2)

(3)

Copyright

The publishers will keep this document online on the Internet – or its possible replacement – for a period of 25 years starting from the date of publication barring exceptional circumstances.

The online availability of the document implies permanent permission for anyone to read, to download, or to print out single copies for his/hers own use and to use it unchanged for non-commercial research and educational purpose. Subsequent transfers of copyright cannot revoke this permission. All other uses of the document are conditional upon the consent of the copyright owner. The publisher has taken technical and administrative measures to assure authenticity, security and accessibility.

According to intellectual property law the author has the right to be mentioned when his/her work is accessed as described above and to be protected against infringement.

For additional information about the Linköping University Electronic Press and its procedures for publication and for assurance of document integrity, please refer to its www home page: http://www.ep.liu.se/.

(4)

3.1.1. TEMPORAL SELECTIONS ... 10 3.2. FOCUS SPECIES ... 11 3.2.1. VITELLARIA PARADOXA ... 12 3.2.2. LANNEA MICROCARPA ... 13 3.2.3. PARKIA BIGLOBOSA ... 13 3.2.4. MANGIFERA INDICA ... 14 3.2.5. AZADIRACHTA INDICA ... 14 3.2.6. KHAYA SENEGALENSIS ... 15 3.3. FIELD DATA ... 15 3.4. SATELLITE DATA ... 17 3.4.1. ETHICAL CONSIDERATIONS ... 17 3.5. CLASSIFICATION METHOD ... 18

3.5.1. DELINEATION OF TREE SEGMENTS ... 18

3.5.2. SPECTRAL PROFILE – TRAINING DATA ... 18

3.5.3. DATA BALANCING ... 19

3.6. CLASSIFICATION MODEL - RANDOM FOREST ... 20

3.7. ACCURACY ASSESSMENT ... 20

3.7.1. MEASURES OF CLASSIFICATION ACCURACY ... 21

4 RESULTS ... 22

4.1. SPECTRAL PROFILES OF TREE SPECIES ... 22

4.2. INTERNAL SPECIES VARIABILITY ... 25

4.3. CHANGE BETWEEN GROWING PERIODS ... 27

4.4. CLASSIFICATION ACCURACY OF EACH IMAGE COMBINATION ... 28

5 DISCUSSION ... 30

5.1. COMPARING IMAGERY ... 30

5.1.1. GROWING PERIODS ... 30

5.1.2. OVERALL CLASSIFICATION ACCURACY ... 31

5.2. SPECTRAL SEPERABILITY OF FOCUS SPECIES ... 32

5.3. CLASSIFICATION PERFORMANCE OF TREE SPECIES ... 32

5.4. FIELD DATA AND DELINEATION QUALITY ... 33

(5)

TABLE OF TABLES

Table 1 - Specification Summary of the Pleiades 1A and 1B Bands and Spatial Resolution

(Gleyzes, Perret and Kubik, 2012). ... 6

Table 2 - Cost (USD) per square kilometre of different satellite images (LAND INFO, 2018). 7 Table 3 – The Ground Truthed Data Pre-Processing Steps and Description of Record Anomlaies ... 16

Table 4 – Statistical features extracted from each tree segment included in the training data ... 18

Table 5 - Applied selection criteria prior to 2017 and 2018 polygon layers. ... 20

Table 6 – Interpretation of Cohen’s kappa values (McHugh M. L. 2012) ... 21

Table 7 - Mean values of reflectance for each species and band. The highest and lowest mean for each band are marked in green and red respectively. ... 22

Table 8 – Maximum (MAX) and minimum (MIN) mean (X̅) reflectance values for each species and band. The highest maximum and lowest minimum for each band is marked in green and red respectively. ... 25

Table 9 - Confusion matrices and statistical metrics for each of the three datasets Correct validation outcomes are marked in green. The highest and lowest producer and users’ accuracy for each dataset is marked in blue and red respectively.. ... 29

Table 10 - Overall accuracies compared across different satellites, bands and number of species ... 31

TABLE OF FIGURES

Figure 1 - Location of study area ... 4

Figure 2 - Pleiades 1a satellite - a few days prior to launch (Gleyzes, Perret and Kubik, 2012) ... 6

Figure 3 - Mean monthly climatic records Saponé between 1982 and 2012 (Climate-Data.org, 2012). ... 9

Figure 4 – Historical Rainfall Records Saponé 2009/2018 from field data records - (M. Karlsson, personal communication, April 1, 2020)... 9

Figure 5 – 2017 & 2018 Wet Season Rainfall compared to the Saponé average between 1982 and 2012 (Climate-Data.org, 2012). ... 10

Figure 6 - Comparison of EOW (Oct 2017) and EOD (May 2018) pan-sharpened subsets of Pleiades satellite images within the study area. ...11

Figure 7 – Images of the focus species selected for this study ... 12

Figure 8 - Photograph showing leaf structure of Vitellaria paradoxa ... 13

Figure 9 - Photograph showing leaf structure of Lannea microcarpa ... 13

Figure 10 - Photograph showing leaf structure of Parkia biglobosa ... 14

Figure 11 - Photograph showing leaf structure of mangifera indica ... 14

Figure 12 - Photograph showing leaf structure of Azadirachta indica ... 15

Figure 13 - Photograph showing leaf structure of khaya senegalensis (Pitchandikulam Herbarium, 2015) ... 15

Figure 14 - Trees removed from ground truthed data due to being considered as "Light Affected" ... 17

Figure 15 – Image examples of the delineated tree segments for each year and species ... 18

Figure 16 - Mean spectral reflectance of each tree species for two different growing periods (a) EOW and (b) EOD ... 24

(6)

Figure 17 - Box-whisker-plots showing the spectral variability within each of the focus species using the combined (2017&2018) training data ... 26 Figure 18 - Scatter chart showing the variation between the 2018 and 2017 object pixel values (y axis=mean, x axis=standard deviation) ... 27 Figure 19 - Overall classification accuracies and kappa values from the 2017, 2018 and combined datasets. ... 28

(7)

ABSTRACT

As we move towards a digital based society, technology continues to improve. It is important to take advantage of this to inform and facilitate our sustainable development goals in the most cost-effective and time efficient manner. By utilising the best available technologies, not only can time savings be achieved, but scope of works can be dramatically increased, particularly with ecological data collection. This study will focus on collecting ecological data (tree species) using developing modern technologies (satellites) with the aim of reaching classification accuracies comparable with ground truthed (real life) records. The study area is in central Burkina Faso approximately 30km south of the capital and is generally described as an agroforestry parklands area. The region suffers greatly from poverty and many people are heavily dependent on the agricultural sector and subsistence farming. As these agroforestry parklands are so critical to many people’s livelihoods, it is important to assess the natural resources available within them to provide the best food security management for the people.

Tree species locations were overlayed on two satellite images acquired during different stages of the annual growing periods in the agroforestry parklands of the study area. From these images, segmentation of individual tree crowns was done manually and used as the reference data for an object-based classification model, which were assessed for the classification accuracies that can be achieved. Three satellite image scenarios were assessed for classification accuracy, including two single image scenarios and a multi-imagery dataset combining both images.

Results indicate that combined images perform the best in terms of overall classification accuracies, closely followed by the end of the wet season growing period. The image acquisition from the end of the dry season was quite poor in comparison, having an overall classification accuracy more than 10% lower than the other scenarios.

Of the focus species assessed in this study, Azadirachta Indica was the clear loser in terms of the number of correctly classified individuals from each model scenario. All other focus species were relatively well classified achieving close to or above 60% accuracies in the multi-imagery classification scenario.

Keywords: Remote Sensing; Burkina Faso; Pleiades Satellite; Tree species; Agroforestry

Abbreviations and Acronyms

EOD End of Dry EOW End of Wet

NDVI Normalised Difference Vegetation Index NIR Near Infrared RF Random Forest RS Remote Sensing

(8)

1 INTRODUCTION

1.1. TECHNOLOGY, LIVELIHOODS AND SUSTAINABILITY

Accurately mapping tree cover is a labour-intensive exercise traditionally, remote sensing (RS) technologies have the potential to change this. RS involves recording measurements of a feature from a distance, without physical contact, typically using an electronic device. Studies have shown that RS can outperform traditional direct measurement field surveys, in terms of speed and scope but not yet in accuracy. Achieving classification accuracies comparable with a direct measurement approach is the main factor limiting the widespread use of RS to classify tree species. For RS to reach these comparable classification accuracies, very high-resolution data is required. The most common RS data collection systems are through Unmanned aerial vehicles (UAV) and satellites. This thesis focuses on satellite data from a 10 x 10 km rural agroforestry landscape area located the West African nation of Burkina Faso. Satellite information was considered the most practical method for obtaining the required RS data, given the remoteness of the location, low availability and high cost of UAVs. As of 16 January 2019 there were 1,957 operational satellites orbiting the Earth reported by the United Nations Office for Outer Space Affairs (UNOOSA, 2019). The vast majority of these cannot provide resolutions required to accurately classify tree species. Amongst the highest publicly available spatial resolution is the data obtained from the Pleiades satellite constellation (Wang, Wang and Liu, 2018). In comparison with satellites that have been used in previous tree classification studies including the Sentinel-2 (10m spatial resolution), Pleiades has the potential to provide significantly higher classification accuracies.

Tree classification using RS data requires measuring light reflectance values at different spectral wavelengths to identify statistical patterns for each tree species. In practise this involves recording RS data and comparing this with unique characteristics of various tree species. Studies focusing on this have increased at rates higher than the general baseline of publication activity from 1980-2015, particularly after 2005 increasing almost exponentially (Fassnacht et al., 2016). This mirrors continuing improvements in satellite data collection, with high resolution imagery leading to increases in tree classification accuracies and consequently additional studies.

Distinguishing and classifying individual tree species through spectral information has been a difficult process. Typically spectral resolutions have not been high enough to allow for species specific characterisation (Wolter et al., 1995). Additionally, the reflectance of tree foliage (leaves, branches and sticks etc) can often have a wider variation than reflectance of surrounding features (bare ground, house rooves, rocky outcrops etc), making them hard to distinguish. As spatial resolution is increased, so too increases the ability to accurately classify individual trees (Wang, Wang and Liu, 2018). For example a study found classification accuracies using Very-High-Resolution (VHR) unmanned aerial vehicle (UAV) images can achieve F-scores (a measure of classification accuracies) reaching 98.2% for the Osmanthus fragrans tree species in a forest nursery landscape (Huang, Li and Chen, 2018). This reinforces the resolution correlation with accuracy, showing that tree classification can reach accuracies comparable with direct observation results.

The focus of this study surrounds an agroforestry parkland landscape within central Burkina Faso. Landscapes like these are widespread throughout West Africa and provide livelihoods for many

(9)

available for the area that the trees can provide. Knowing this information can be used to predict the productivity of parklands in terms of economic value, human and ecological health. It is also vital to understand the effects of future changes on parklands, such as climate change, water shortages or intensified land use. These are key challenges that could impact the health and abundance of trees, affecting livelihoods of subsistence farmers who depend on them (Ræbild, B. Hansen and Kambou, 2012).

SUMMARY AND OBJECTIVES

The aim of this study is to find out what levels of accuracy can be expected when classifying the dominant tree species of a selected study area in the Saponé region, Burkina Faso using Pleiades satellite imagery. This study has the potential to fill information gaps and add knowledge regarding the ability to use RS data from Pleiades to classify trees in the study area.

Pleiades satellites high-resolution imagery was used to distinguish individual trees species based on their spectral properties. The information was recorded by the reflectance sensors mounted to the Pleiades satellites. Two different growing periods of an annual cycle were compared, namely the end of the wet season in October 2017 and the end of the dry season in May 2018. The results from this study could help inform management strategies regarding abundance of tree resources and the status of food security in the region. Thus, the results of this thesis are relevant for work around the Sustainable Development Goals, particularly goal no. 2, Zero Hunger (United Nations, 2016) and could be used by Food and Agriculture Organization of the United Nations (FAO), International Fund for Agricultural Development (IFAD), FAO's Monitoring African Food and Agriculture Policies (MAFAP) and or the World Bank Group.

RESEARCH QUESTIONS

The research questions which form the focus of this thesis are as follows: Using Pleiades satellite imagery;

• What growing periods perform better for classification of parkland tree species? o End of wet season (EOW) October; or

o End of dry season (EOD) May.

• What Wester African parkland tree species are most easily classified using Pleiades satellite imagery?

2 BACKGROUND

2.1. BURKINA FASO

Burkina Faso is a country gripped by poverty. In the last decade Burkina Faso’s Human Development Index (HDI) value has consistently ranked amongst the lowest in the world (ranks 185 out of 188 countries) (World Food Program USA, 2019) and approximately 40 percent of its population lives below the poverty line (World Bank, 2018). Agricultural production is critical to the national economy with close to 80% of the active population employed in the sector (The World Bank, 2019) and food security is still a key issue. Over the last ten years the population has increased at over 3 percent per annum across the country, and 7 percent per annum in the capital city, Ouagadougou adding to the food security pressure (World Bank, 2018).

2.1.1. LOCATION STUDY AREA – SAPONÉ

The study area is located in the rural commune of Saponé (12o_{03’ N and 1}o_{43’ W altitude 200 m),}

(10)

comprised of parklands, woodlands, fallow land and settlements (Karlson, 2015). The tree cover is dominated by the species Vitellaria paradoxa, Parkia biglobosa, Lannea microcarpa and

Mangifera indica (Karlson, Reese and Ostwald, 2014), and the understory consists of grasses,

shrubs and crops.

Figure 1 - Location of study area

2.1.2. LAND USE – AGROFORESTRY PARKLANDS

The dominant land use in the study area is subsistence farming over “agroforestry parklands”, a highly sustainable form of farming that has been practiced for thousands of years in West Africa (Boffa and FAO, 1999). These parklands are the most common farming systems found in West Africa and are managed for crops, livestock feed, wood, food and nonwoody tree products including medicine (Boffa and FAO, 1999) (The Forests Dialogue, 2011). Forming these parklands involves selective clearing of greenfield (untouched) landscapes to allow for crop production, often with certain trees being protected due to their respective food and medicinal values (Boffa and FAO, 1999). Crops grown in the study area are mainly cereals, such as sorghum, millet and maize, and legumes, including cow pea and peanut (Nikiema, 2005).

There are several methods through which these parklands are formed, broadly defined by the amount of human intervention that has occurred during their formation. The least modified of these include transient “residual” parklands, where tree species are left partly based on their usefulness but also partly on how hard they are to remove. That is if they are too hard to clear, they are often just left in place. So the trees remaining in this type of parklands are not solely a reflection of their importance to the farmers resulting in a relatively high species richness (Boffa and FAO, 1999). A high native species richness is likely to correlate with a higher provision of ecosystem health services. Another way parklands can be formed is through a more indiscriminate clearing process. Forming parklands through clearing often results in the success of tree species which can regenerate the fastest and survive with little shade. These species end up dominating the cleared parklands. Understanding the dominant tree species of an operational agroforestry parkland can give an insight into the formation methods. Knowing how parklands were formed can be useful

(11)

2.2. REMOTE SENSING

Everything reflects light and energy in different ways enabling us to make visual sense of objects (NASA, 2006). To assess the best method of tree species classification we need to look at the reflectance of many wavelengths which include both the visible and non-visible spectrums. RS is one such method which allows us to obtain this information with relative ease. RS can obtain information about an object without having to make direct physical contact with it. This is done via electronic sensors which receive electromagnetic radiation reflections from the object. The platform which these sensors are mounted can be on satellites, aircraft or UAVs. The information is recorded into separate spectral bands, which can be represented visually in images as individual pixels. The benefits of RS include being able to observe large areas in a short period of time, repeat observations over time to create time series data, focus on specific electromagnetic spectrums and obtain the information without having to visit the area (Forster, 1985).

With the invention of the first known cameras, so to was remote sensing born. In the 1840s, the first known records of remotes sensing were collected with cameras mounted to hot air balloons. Across both world wars, aerial recon information became invaluable and this pressure drove further advances in technology. Probably the most significant leap in the fields history occurred in the 1960s and 1970s where radar imaging systems were sent into the earths orbits (otherwise known as satellites) to record and send information back to earth. From the cost of sending a hot air balloon to record and photograph an area of land, to today’s modern satellites, the cost of producing images has consistently been reduced.

2.2.1. SATELLITE INFORMATION - PLEIADES

The Pleiades satellite constellation provided the images used for this study. Built under management of the French-Italian Optical and Radar Federated Earth Observation (ORFEO) programme, the constellation is made up of two satellites, Pleiades 1A and Pleiades 1B, which were launched on December 17, 2011 and December 2, 2012 respectively (Artigues, Greslou and Baillarin, 2017). Both satellites (1A and 1B) have a sun-synchronous orbit of 695 km altitude, swath width of 20km and otherwise the same technical specifications (Gleyzes, Perret and Kubik, 2012). They have a continuous orbit with a daily revisit capacity to deliver high resolution optical images of any point on Earth (Gleyzes, Perret and Kubik, 2012). Pleiades 1A is shown in Figure 2.

(12)

Figure 2 - Pleiades 1a satellite - a few days prior to launch (Gleyzes, Perret and Kubik, 2012)

Sunlight provides the reflections which are received by time delay integration (TDI) charge-coupled device (CCD) sensors mounted to the Pleiades satellites. The TDI-CCD image sensors are used for the panchromatic band as they account for object movement providing increased sensitivity when recording fast moving objects (Hamamatsu, 2016). The multispectral information is received by five additional sensors, each with four lines of assembly corresponding to four spectral bands blue, green, red and near infrared (NIR) (Gleyzes, Perret and Kubik, 2012).

SPATIAL RESOLUTION

When satellite data is relayed into images, spatial resolution generally refers to the size of the image pixels. The smaller the pixel size the higher the image resolution. Pleiades imagery has a spatial resolution of 2 m for the multispectral bands (blue, green, red, NIR) and 0.5 m for the panchromatic band as presented in Table 1 (Gleyzes, Perret and Kubik, 2012).

Table 1 - Specification Summary of the Pleiades 1A and 1B Bands and Spatial Resolution (Gleyzes, Perret and Kubik, 2012).

Band Name Band Wavelength (µm) Spatial Resolution (m) Band Reference #

(used for this study)

Panchromatic 0.48 – 0.83 0.5 NA

Blue (Visible) 0.43 – 0.55 2 1

Green (Visible) 0.49 – 0.61 2 2

Red (Visible) 0.60 – 0.72 2 3

Near infrared (NIR) 0.75 – 0.95 2 4

Pixel size is an important factor when mapping land cover. Typically, higher resolution of data increases accuracy of tree species classification (Boyd and Danson, 2005). However the trade-off for this increasing resolution accuracy usually comes with higher costs, large datasets, processing times and less reliability of covered areas (Boyd and Danson, 2005). A summary of available images and costs through various collection methods is presented in Table 2.

(13)

Table 2 - Cost (USD) per square kilometre of different satellite images (LAND INFO, 2018).

Product WorldView-_{1, 2 &3} QuickBird _(60cm) GeoEye-₁ IKONOS _(80cm) Pléiades _1A/1B

High Resolution Archive Pricing

(50cm) 14 14 14 10 12.5

When compared with UAV image collection, satellites are a much cheaper option. One commercial service operating out of Canada can provide image collection at a rate of $3 (USD) per acre, which works out to be approximately $750 per square kilometre (Deveron, 2020).

2.2.2. LEAF MORPHOLOGY AND REFLECTANCE

Leaf sunlight reflectance is inherently related to the structure, pigments and physiological function of the tree species and the leaf itself. Surface reflectance of the leaves is the dominant factor in measuring spectral variation through remote sensing in most tree species (Sims and Gamon, 2002). Electromagnetic radiation (sunlight) is absorbed differently according to the various physiological properties of the tree. Leaf structure and chemical composition (chlorophyll, water and cellulose) dictate the spectral interaction (Tzionas, Papadakis and Manolakis, 2005). Leaf morphology affects how light is scattered and ultimately reflected to the satellite. This unique absorption and reflection pattern is called a spectral signature. Differences in the reflectance patterns can be used to distinguish tree species. Biological processes, such as leaf shedding and flower generation, can also be used as a predictor of tree species as these processes results in different reflectance (Avuclu and Elen, 2017). Commonly, leaf reflectance values are higher for the NIR wavelengths compared to the visible spectrums (Sims and Gamon, 2002). Tree crown reflectance can also be affected by tree density, proximity to neighbouring trees and influence of surrounding features, such as shadow or bare soil.

As reflectance mirrors changes to phenological attributes of trees, losing leaves or shrinking in size over dryer periods will result in varying reflectance values. Satellite images taken from periods where tree phenology varies the greatest is predicted to have the greatest impact on classification accuracy (Sheeren et al., 2016).

2.3. CLASSIFICATION APPROACHES

The process of RS image classification involves allocating a defined class or attribute to each individual pixel of which the images are composed. This is done by converting the spectral information obtained by the sensors into user defined information classes (such as tree species, land use type or landscape feature). There are several ways of allocating information classes to pixel units, but the two most popular methods are pixel-based and object-based classification (Zerrouki and Bouchaffra, 2014).

2.3.1. PIXEL-BASED

Pixel-based classification involves, in its simplest, using spectral information as the basis for predicting information classes of each pixels across an image one pixel at a time (Weng, 2012). In this approach, each pixel’s spectral information is used as an input into a classification algorithm which would predict the classification for that individual pixel. A fundamental limitation of pixel-based classification is that information from surrounding pixels, which may help in correctly identifying the target pixel’s class, is not used. For example, in a cluster of pixels that comprise of an object, each pixel would be treated as a separate entity. Thus, a case for using object-based

(14)

approaches can be made when dealing with individual tree crowns, especially when the spatial resolution of the satellite imagery used is high.

2.3.2. OBJECT-BASED

In the past, RS data may have contained resolutions where one pixel may have previously generalised an entire feature (such as a tree), now pixels are at such a resolution where many may be required to form part of each feature (Jónsson, 2015). With increasing resolution of RS data, spectral information can be obtained from much smaller areas and object-based classification methods have been developed to utilise these higher spatial resolutions. Object-based image classification starts with segmenting an image into meaningful shapes, sizes or features which represent real-world objects such as buildings, lakes or trees (Dey, Zhang and Zhong, 2010). The segmentation of the image and grouping of pixels form the basis for the term “object”. Each object contains spectral, textural and contextual information which can be summarised by statistical features such as the maximum, mean, standard deviation and area. These statistical features can be used to determine the likelihood of each pixel within each object being classified as the correct real-world feature. In this study, the tree crowns are the objects defined by image segmentation, and all pixels within the object are attributed to in situ species information. Shape characteristics and spatial references also differentiate object-based classification from the pixel-based approach (Zerrouki and Bouchaffra, 2014).

Studies have found object-based approaches in most cases outperform pixel-based approaches in terms of classification accuracy, but spatial resolution of the RS data was a determining factor (Hussain et al., 2013) (Duro, Franklin and Dubé, 2012) (Zerrouki and Bouchaffra, 2014). The performance difference is most significant where the spatial resolution of the pixel sizes are smaller than the targeted objects for classification (Ng et al., 2017). Using the resolution of today’s modern satellites, spectral variability within a tree crown is better accounted for with object-based compared to pixel based classification (Immitzer, Atzberger and Koukal, 2012).

One drawback of object-based classification, however, is that under-segmentation and over segmentation are potential shortcomings (Zerrouki and Bouchaffra, 2014). This occurs when there are errors in accuracy of segmentation boundaries and they do not represent the real-world objects which leads to false statistical feature summaries and lower classification accuracies (Liu and Xia, 2010).

3 MATERIALS AND METHODS

3.1. STUDY AREA CLIMATE

Saponé falls within the Sudan-Sahel climatic zone of West Africa (Le Houerou, 1990), a region that has experienced a dramatic change in climate over the past 30–40 years with declining annual rainfall (Mertz et al., 2012). From 1982 to 2012 the region has a recorded mean annual temperature of 27.6°C and mean annual rainfall of 794 mm (Climate-Data.org, 2012). The mean monthly rainfall and temperature summary is presented Figure 3.

(15)

Figure 3 - Mean monthly climatic records Saponé between 1982 and 2012 (Climate-Data.org, 2012).

There is a distinct dry season which spans from November to March in the study area. This is followed by a period of intermittent rainfall beginning in April / May which leads into the dominant rainy (wet) season June to October (Bayala, Roméo and Sanou, 2013). The wet season is characterised by heavy rainfall and lower temperatures (Ilstedt et al., 2016). The mean annual rainfall can vary greatly from year to year as shown in Figure 4.

Figure 4 – Historical Rainfall Records Saponé 2009/2018 from field data records - (M. Karlsson, personal communication, April 1, 2020)

1221 891.2 1011 850.57 568.6 566.9 926.3 1008.4 852.7 789.1 0 10 20 30 40 50 60 70 0 200 400 600 800 1000 1200 1400 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 N u m b er o f d ay s with ra in TOT AL YE ARLY RAIN FALL m m

(16)

The total annual rainfall falls almost exclusively between the months of April to October as shown in Figure 5.

Figure 5 – 2017 & 2018 Wet Season Rainfall compared to the Saponé average between 1982 and 2012 (Climate-Data.org, 2012).

The 2017 wet season had slightly higher recorded rainfall than average, and the 2018 Dry season lasted slightly longer than average (Figure 5). This should be ideal conditions for detecting any phenological changes induced by water availability between the two growing periods.

3.1.1. TEMPORAL SELECTIONS

Satellite images from two annual growing periods were compared in this study. These were October 2017 and May 2018 corresponding to the EOW and EOD respectively. The images were acquired from two vastly different growing periods to account for any phenological changes. The images would be used individually, and in combination, for tree species classification. A snapshot of both the EOW and EOD images is presented in Figure 6.

(17)

Figure 6 - Comparison of EOW (Oct 2017) and EOD (May 2018) pan-sharpened subsets of Pleiades satellite images within the study area.

3.2. FOCUS SPECIES

The two most common tree species in the study area are Vitellaria paradoxa (Vit) and Parkia

biglobosa (Par) (Bayala et al., 2006) (Karlson, 2015). Other prevalent species in the area include Lannea microcarpa (Lan), Mangifera indica (Man), Azadirachta indica (Aza) and Khaya senegalensis (Kha). Together, these six make up the focus species in this study and represent the

majority (88.4%) of the total reference samples collected. Images of each are presented in Figure 7.

Adansonia digitata and Faidherbia albida (Winter thorn) are other species present in the

agroforestry areas, but not focussed on in this study (Bayala et al., 2006) (Dal, 2016).These tree species share a similar characteristic with the focus species, that is that they are highly useful to the landowners and are selectively retained by many cultivation farmers when clearing land (Nikiema, 2005). Mangifera indica, which is not native, is planted either as individual trees or as within designated orchards for their use to provide fruit. Azadirachta indica is also a non-native but grows well in the study area and was introduced (planted) for its many benefits, including as a source of medicine, food, fuel and shade (Orwa, 2009).

(18)

1 - Vitellaria paradoxa 2 - Lannea microcarpa

3 - Parkia biglobosa 4 - Mangifera indica

5 - Azadirachta indica 6 - Khaya senegalensis

Figure 7 – Images of the focus species selected for this study

3.2.1. VITELLARIA PARADOXA

Commonly known as the Shea tree, this species is native to the Sudanian savanna zone. It is deciduous and can grow to about 25 meters in height. The leaves have an oblong shape are 10–25 cm long and 4–14 cm wide and concentrated towards the ends of branches Figure 8. In the agroforestry parklands of the study area, Shea trees are rarely leafless (Bazié, 2013). They produce round fruits that contain the nuts required to make highly valuable shea butter and vegetable fat. Only palm oil is more important in terms of economic value in Africa (Bayala et al., 2006). Shea butter has many uses including medicinal, cosmetic and in cooking (Hall et al., 1996). The tree also provides valuable ecological benefits including the conservation of soil and water resources. It is common for unhealthy trees in the area to be cut down for timber and fuel wood (Moore, 2008).

(19)

Figure 8 - Photograph showing leaf structure of Vitellaria paradoxa

3.2.2. LANNEA MICROCARPA

Commonly known as the African grape, it is native in the study area, deciduous and can grow up to 16 meters high. It has a clustered leaf structure which grow at the end of branches and create a dense tree crown (Fern, 2019b). Leaflets are ovate in shape, approximately 5–13 cm long × 2.5–6 cm wide and can be seen in Figure 9 (Marquet and Jansen, 2005). Leaves often fall off early into the dry season and regrow again at the beginning of the wet season. Bark of the African grape can be used as a red-brown coloured dye for textiles and other materials. The tree also provides benefits, including medicinal, food, wood fuel and is a good source of shade for livestock (Ganaba, Ouadba and Bognounou, 1999).

Figure 9 - Photograph showing leaf structure of Lannea microcarpa

3.2.3. PARKIA BIGLOBOSA

Commonly known as the African locust bean, it is native to the study area, is not deciduous and grows to heights of 15-20m at full maturity. The leaves are 30-40 cm long and bipinnately compound as shown in Figure 10. The tree crown is dense, wide spreading and generally circular in shape. Many parts of the tree (seeds, fruit, flower buds and leaves) are used as food sources for people and livestock in the area (Ræbild, B. Hansen and Kambou, 2012). Environmental benefits are also provided by the African locust bean, such as protection from soil erosion, can act as a windbreaker and indirectly helps to fertilise the soil through the animal droppings that occur while they congregate underneath for shade (Hopkins, 2008).

(20)

Figure 10 - Photograph showing leaf structure of Parkia biglobosa

3.2.4. MANGIFERA INDICA

Commonly known as the Mango tree, this species is one of the oldest cultivated plants in the world having been used first for this purpose in India as much as 4000 years ago (Shah et al., 2010). It is an evergreen tree and grows to approximately 20m in height at full maturity in the study area. The leaves are simple narrow oval shape tapering toward each end and can grow up to 30cm long and 8cm wide Figure 11 (Fern, 2019c). Mango has been introduced to the study area mainly for the highly sought after fruit, which is one of the most popular in the world (Shah et al., 2010). It is the most common fruit crop in Burkina Faso and is an important source of income and subsistence for many farmers. A study reported that in 2012 about 14 000 hectares in Burkina Faso was occupied by the mango tree. As a percentage of the surface of all perennial crops this represented 58% and thus it is a very important tree to the national economy (Vayssières et al., 2012).

Figure 11 - Photograph showing leaf structure of mangifera indica

3.2.5. AZADIRACHTA INDICA

Commonly known as the Neem tree, Azadirachta Indica is a fast-growing evergreen that can grow up to 30 meters tall, is drought tolerant and can survive with as little as 130 mm of rainfall per year (Gaméné et al., 1996). It is native to the Indian subcontinent region and has been introduced (planted) to the study area for its many economic and environmental benefits (Orwa, 2009) (Kumar and Navaratnam, 2013). Leaves are simple pennate (resembling a feather) which alternate along the tree stem and concentrate towards the end of branches and can range between 20-40cm long at maturity Figure 12. Early leaves can be more red / purple in colour from mature leaves and in extreme drought the tree may shed its leaves (BioNET-EAFRINET, 2011). This species is used to help restore degraded lands in drier climates as establishes itself quickly with little rainfall so helps to minimise soil erosion, improve soil fertility and increase the water holding capacity of the underlying soils (Kumar and Navaratnam, 2013). It is also highly valuable for medicinal purposes (Neem Research, 2019).

(21)

Figure 12 - Photograph showing leaf structure of Azadirachta indica

3.2.6. KHAYA SENEGALENSIS

Commonly known as the African mahogany, this is a large tree (in both height and tree crown), growing up to 30 m tall and spreading up to 16m in diameter (Fern, 2019a). Usually evergreen it can shed its leaves in periods of drought or harsh climates. It is native to central and western parts of Africa, including some regions in Burkina Faso (Arnold et al., 2004). The leaves cluster toward the edge of branches, are pennate in shape and vary between 7-15 cm long and 3-4 cm wide (Pitchandikulam Herbarium, 2015) Figure 13.

Figure 13 - Photograph showing leaf structure of khaya senegalensis (Pitchandikulam Herbarium, 2015)

The timber of the African mahogany is highly sought after and as a consequence the species has been widely exploited (Plantnet-project, 2017). In 1998 the tree was classified as vulnerable on the IUCN Red List of Threatened Species, where it still remains today (World Conservation Monitoring Centre, 1998). Medicines can also be obtained from the seeds and leaves of the plant (Fern, 2019a).

3.3. FIELD DATA

Field reference information was collected from the study area during two field campaigns, one from 2012 and another from 2017. During these campaigns, tree species were physically identified by a local botanist and spatially referenced using a global positioning system (GPS) device (Karlson, 2015). The 2012 campaign involved random sampling of stratified areas, selected to account for the range of vegetation types and tree densities found in the study area. Stratification involved splitting the study area into four classification categories based on normalized difference vegetation index (NDVI) representing percent tree canopy cover (TCC). Large plots (each 50 × 50 m) were distributed across areas classified as Low, Medium or High TCC (Karlson, 2015). Data was collected from trees with a ≥ 5 cm diameter at breast height (DBH) included georeferenced

(22)

locations, crown size, species type and tree height (Karlson, 2015). The 2017 field campaign followed a similar structure to that undertaken in 2012, however only areas classified as agricultural fields were sampled. The field reference information used in this thesis is the species type and location, herein known as “ground truthed” data.

Ground truthed data was comprised of 1202 spatially referenced tree records, containing 29 unique species. The records were imported into ArcGIS Pro and georeferenced to the EOW and EOD images. From this initial import several tree records which obviously deviated from what was standard or expected were removed. This occurred over several rounds of data processing, each occurring at a new step of the project as the scope of the thesis took shape. Ultimately four rounds of data quality processing were conducted and a final count of 802 records were selected for analysis as shown in Table 3.

Table 3 – The Ground Truthed Data Pre-Processing Steps and Description of Record Anomalies

Anomaly Anomaly Count Removed from ground truthed records

Point not spatially referenced 1 Round 1

Tree not visible (compared to Oct 2017) 11 Round 1

Tree location unclear 25 Round 2

Tree part of mango plantation 22 Round 2

Affected by external light source 222 Round 3

Statistical outlier 23 Round 4

Filtered out non-focus species 96

RECORDS SELECTED FOR ANALYSIS 802 Not Removed

Total 1202

ANOMOLY DESCRIPTIONS

In the first round of data processing, we identified and removed one tree record that was not spatially referenced. In addition, there were 11 records where a tree was visible in the 2017 aerial image but not observed, or “missing”, in the 2018 image. This may have been because the trees were cut down, caught fire or otherwise destroyed during the period between the image acquisition dates.

During the second round, it was clear several tree records were from areas of mango plantations. This study aims to classify specific trees within agroforestry landscapes and for this reason records of trees located within areas of mango plantations were removed, most of these species were Man. There were also several records where it was not clear which trees the data point was attributed to and these were removed.

The third round of processing involved visually identifying tree locations that appeared to be influenced from land with abnormally high light reflection. This round occurred after polygons had been generated around the tree crowns for each image (tree segmentation), to see if any external light had extended into the segmented tree crown. If there was an obvious influence from an external reflection such as bare ground, soil or buildings that had encroached into the tree segmentation from either of the image dates, then these records were labelled as “Light Affected”

(23)

Figure 14 - Trees removed from ground truthed data due to being considered as "Light Affected"

Round four involved identifying outliers, which were then revisited and redrawn in cases that required. The outliers were identified by comparing the z-scores calculated for each tree segment based on the year, reflectance band and species combination. Z scores for each combination were filtered for values above 2.5 for both the mean and standard deviation. The result was ~150 records required re-visiting to ensure that each tree segment was drawn accurately. Of these ~150 records, in many cases the tree segmentation just needed re-shaping and remained in the ground truthed data, but 23 trees were considered as an extremely poor fit with their respective species groupings and were removed. Lastly, to streamline next stage of classification, we kept only species which had more than 20 individual tree records known as the focus species. After all rounds of processing the ground truthed data were complete, 802 records remained.

3.4. SATELLITE DATA

Multispectral satellite images were obtained from the Pleiades constellation in October 2017 (EOW) and May 2018 (EOD) respectively. The images both covered the same 10 x 10 km region over the study area. Pre-processing of the Pleiades satellite images involved georeferencing the data to a WorldView-2 image from 2012 and calibration to top-of-atmosphere reflectance using Orfeo Toolbox (OTB) a library for remote sensing image processing. Monteverdi is the graphical user interface which was used to process the metadata file and transform the data into useable, Pleiades satellites derived images where the pixel values represent reflectance (Herbert J. Kramer, 2002). Pan-sharpening of the images was required to fuse the multispectral bands with the panchromatic band to increase the spatial resolution (Alparone et al., 2007). After pan-sharpening the resulting multispectral images had a 0.5 m spatial resolution, meaning that each pixel in the image was 0.5 x 0.5m in dimensions. The two images were imported to ESRI’s ArcGIS Pro software program (version 2.3.1) as a raster file and projected to the WGS 1984 UTM coordinate system (Zone 30N).

3.4.1. ETHICAL CONSIDERATIONS

There are no obvious major ethical concerns with the scope of this study. Some ethical considerations for remote sensing in general are that policy development often lags behind advancements in technology resulting in access to sensitive data before policies and regulations are put in place (Slonecker, Shaw and Lillesand, 1998).

(24)

3.5. CLASSIFICATION METHOD

3.5.1. DELINEATION OF TREE SEGMENTS

Delineation, or the process of drawing a digitalised line around each tree crown, has a direct influence on the outcomes of the classification models (Immitzer, Atzberger and Koukal, 2012) (Fassnacht et al., 2016). Delineation can be performed manually or automated via algorithms. Automated methods often build on previous studies and require a significant time period of design and are arguably a field of research in its own right (Fassnacht et al., 2016). Therefore, manual delineation was selected for this study.

Manual delineation involved drawing polygons around each tree crown (n=802) to generate a “tree segment”. The tree segments are estimates of the tree’s representative pixel composition. Shaded areas and bare ground were selectively excluded to ensure the tree segments were representative of the tree foliage itself. The tree segments are, in general, proportionally smaller than the actual tree size. This is consistent across all species and therefore are comparable with themselves in terms of relative tree size. A visual representation of delineated tree segments are shown in Figure 15.

Aza Kha Lan Man Par Vit

2017

2018

Figure 15 – Image examples of the delineated tree segments for each year and species

3.5.2. SPECTRAL PROFILE – TRAINING DATA

Training data refers to the statistical data used in the classification model. Features which were selected to form the training data are summarised in Table 4. These features were selected based on their success in previous studies (Roth et al., 2019). The information was extracted using the ArcGIS Pro zonal statistics tool.

Table 4 – Statistical features extracted from each tree segment included in the training data

Feature Type # of Feature Name

Spectral Characteristic 2 Mean _{Standard Deviation}

Pleiades Spectral Bands 4

Blue Green Red NIR

(25)

Spectral and textural characteristics of each tree segment were extracted from the satellite images. This included the reflectance of each tree segment (mean and standard deviation) across each spectral band and the area. The normalized difference vegetation index (NDVI) of each tree segment was also included as training data, calculated as per the following equation.

𝑁𝐷𝑉𝐼 = (𝐵𝑎𝑛𝑑 4 (𝑁𝐼𝑅) − 𝐵𝑎𝑛𝑑 3 (𝑅𝑒𝑑)

𝑁𝐼𝑅 + 𝑅𝑒𝑑 )

In total there were 22 predictor variables for each individual tree segment forming the training data. The predictor variables were a combination of each feature is as follows:

• 2 (spectral characteristic) x 5 (spectral bands & NDVI) x 2 (for each images) = 20; plus • 1 (spatial attribute) x 2 (for each images) = 2

The training data was structured based on three scenarios: • EOW single image;

• EOD single image; and

• EOW & EOD combined images (Combined).

These months were chosen as the corresponding growing periods have not previously been assessed in the study area.

3.5.3. DATA BALANCING

To allow for suitable classification through the machine learning algorithm discussed in Section 3.4, the training data needed to be balanced. As with most classification algorithms, there is a curse of learning. This causes imbalanced training data tending to prioritise the classification accuracy towards the dominant feature class (Chen, Liaw, & Breiman, 2004). In this case the dominant feature class was Vit with 431 records out of the total 802, or 53.7%, meaning the training data was heavily skewed towards this species. In order to balance this prior to the classification model, tree segments were filtered by a set size criterion for each species as shown in From the initial size filter, it was clear there were still too many Vit relative to the other species (n=227) so an additional filter was placed on this species only. This filter was set to exclude trees which did not occur within nominated agroforestry parkland areas, and as a result we were able to get the total down to an appropriate number for the classification model (n=152).

Table 5.

From the initial size filter, it was clear there were still too many Vit relative to the other species (n=227) so an additional filter was placed on this species only. This filter was set to exclude trees which did not occur within nominated agroforestry parkland areas, and as a result we were able to get the total down to an appropriate number for the classification model (n=152).

(26)

Table 5 - Applied selection criteria prior to 2017 and 2018 polygon layers.

Focus Species Criteria _(m2₎

“FINAL SELECTION” based on size criteria Average delineation size m2_{of 2018} Average delineation size m2_{of 2017} Vitellaria paradoxa >30 152* 40.01 48.04 Lannea microcarpa >15 95 31.16 40.48 Parkia biglobosa >30 108 129.59 129.13 Mangifera indica >30 69 50.94 58.38 Azadirachta indica >15 33 28.40 41.01 Khaya senegalensis >0 24 180.05 204.08

*Vitellaria also filtered for individuals within nominated agroforestry parkland areas.

3.6. CLASSIFICATION MODEL - RANDOM FOREST

Random Forest (RF) is a supervised machine learning algorithm used to create classification models based on datasets and is one of the most powerful methods used for image classification (Roe, Yang and Zhu, 2006) (Svetnik et al., 2003). RF works by creating multiple decision trees to split datasets based on certain criteria and then merges them together based on their predicted feature class. A decision tree can be thought of as series of questions asked about the dataset which would eventually lead to a prediction of that class (Liaw and Wiener, 2002). In remote sensing studies focussing on tree species, RF is preferred to many other classification models due to its high accuracies with supervised classification, straightforward interpretation and ability to easily parameterise (Roth et al., 2019) (Persson, Lindberg and Reese, 2018). A recent object-based classification study found RF to outperform Support Vector Machine (SVM) and Linear Discriminant Analysis classifiers for of tree species in the City of Tampa, Florida USA (Pu, Landry and Yu, 2018). Li et al also found that using RF was faster and more accurate than both SVM and pixel-based analysis (Li et al., 2015) however this study was classifying forested landslides, not individual tree species.

RF models were then fitted to the three sets of training data (EOW, EOD and Combined), each processed through a separate iteration of the classification model using the R studio (Version 1.1.463) ‘Rattle’ package user interface (Version 5.2.0) (Williams, 2017). Each model was set with the number of decision trees (n=1000) and number of variables (n=4).

3.7. ACCURACY ASSESSMENT

A k-fold cross-validation procedure was used to determine the predictive accuracy of the RF classification models. The procedure divides data into separate portions to test the classification model through resampling (Congalton and Green, 1998). The separated portions are processed through the model separately and used to predict the “unknown” set of records (validation data). k represents the number of folds in which the data has been divided. In this case the training data has been divided into five equal and random distributions. Four portions, or folds, (80%) of this data were used to “form” the algorithm and the remaining one portion (20%) was used for validation of the results. The R-studio ‘rattle’ package was used iterate the resampling procedure and the sequence five times so each portion was used once for validation. The results from the

(27)

Marco Vanetti, which provides different estimates of classification accuracies as discussed below (Vanetti, 2007).

3.7.1. MEASURES OF CLASSIFICATION ACCURACY

There are four estimates of classification accuracies provided by the confusion matrix; overall accuracy (OA), Cohen’s Kappa coefficient (K value), user accuracy and producer accuracy (Congalton and Green, 2008) (Jónsson, 2015). These statistics provide an indication of how well the RF classification model performs. Overall accuracy, expressed as a percentage, is a measure of how many of the predicted class records matched the “ground truthed” reference data, or simply how many trees were accurately classified. This is the most commonly cited statistic for tree species classification using machine learning algorithms, as has been used in many other studies (Persson, Lindberg and Reese, 2018)(Karlson et al., 2016)(Immitzer, Atzberger and Koukal, 2012)(Colgan et al., 2012). The user’s accuracy is a measure used to reflect the likelihood that information from the training data matches the ground truthed records. The producer´s accuracy indicates how well each sample classifies each feature class (Congalton, 2015) (Congalton and Green, 1998). The Kappa statistic (K) is used to reflect how data classifiers assign the same score to the same data item, known as inter-rater reliability (McHugh, 2012). It takes into account the chance that correct classifications could be purely due to chance as well as due to an actual agreement (McHugh, 2012). K values ranges from 0 to 1 with the closer the value reaches to 1 the less likely that classification is due to random agreement as shown in Table 6.

Table 6 – Interpretation of Cohen’s kappa values (McHugh M. L. 2012)

Value of Kappa Level of Agreement % of Data that are Reliable

0 – 0.20 None 0–4%

0.21 – 0.39 Minimal 4–15%

0.40 – 0.59 Weak 15–35%

0.60 – 0.79 Moderate 35–63%

0.80 – 0.90 Strong 64–81%

(28)

4 RESULTS

4.1. SPECTRAL PROFILES OF TREE SPECIES

Average reflectance values for each tree segment, species and band combination are presented in Table 7. These records provide a snapshot overview of the spectral and textural properties for each tree species. The species with highest average standard deviation (sd), or data variability, across all bands was observed within Vit (sd=0.062). Vit also had the greatest number of individual records (n=152). The equal lowest sd were within Aza and Lan (sd=0.047). This is particularly interesting for Lan as this species had a relatively high number of individual records (n=95) and yet still has one of the lowest average standard deviations. This indicates that as a species Lan is quite homogenous in terms of spectral reflectance.

The species with the highest average mean (X̅) reflectance for each species and band combination was Man (X̅ = 1.054), particularly high in the NIR and NDVI bands. This indicates that Man has a higher density of tree leaves and foliage as is peaked with higher spectral wavelengths. The lowest average mean reflectance for each species and band combination was found in Kha (X̅ = 0.987). The lowest average mean reflectance for NDVI was found in Lan (X̅ = 0.333) indicating it has the lease dense tree crown in terms of leaf foliage / vegetation growth.

Table 7 - Mean values of reflectance for each species and band. The highest and lowest mean for each band are marked in green and red respectively.

Species # tree segments Reflectance B1 (Blue) B2 (Green) B3 (Red) B4 (NIR) INDEX (NDVI) Total Aza 33 MEAN 0.125 0.126 0.125 0.286 0.343 1.005 STD 0.006 0.005 0.005 0.017 0.015 0.047 Kha 24 MEAN 0.113 0.121 0.122 0.281 0.350 0.987 STD 0.007 0.007 0.007 0.022 0.017 0.060 Lan 95 MEAN 0.130 0.131 0.128 0.284 0.333 1.007 STD 0.006 0.005 0.005 0.017 0.015 0.047 Man 69 MEAN 0.122 0.130 0.126 0.310 0.365 1.054 STD 0.006 0.006 0.006 0.022 0.017 0.058 Par 108 MEAN 0.118 0.123 0.124 0.280 0.355 0.998 STD 0.005 0.005 0.004 0.017 0.019 0.050 Vit 152 MEAN 0.128 0.130 0.128 0.298 0.357 1.041 STD 0.006 0.006 0.005 0.023 0.022 0.062 Different sections of wavelengths have been “cut and measured” in sections, known as bands. Each of the bands are different and together form the spectral signature of a feature in question. Mean spectral profiles of the focus species for the EOW and EOD scenarios are presented in Figure 16. As expected, NDVI provided the highest mean reflectance across all species and model scenarios, followed closely by the NIR band. The visible spectrums had significantly lower values in comparison.

(29)

range was 0.333 - 0.365 and the band 3 range was 0.122 - 0.128. Across all model scenarios, the species order of highest to lowest reflectance for NDVI was Man, followed by Vit, Par, Kha, Aza and lastly Lan.

The tree species mean reflectance varies significantly between EOW and EOD images in the visible spectrums (band 1 - blue, band 2 - green & band 3 - red), but very little between the higher wavelength spectrum (band 4 – NIR & NDVI). Leaf morphology drives reflectance values in the higher wavelength spectrums (NIR & NDVI) (Persson, Lindberg and Reese, 2018). The similar reflectance values for NIR and particularly NDVI during both images, indicates that the tree foliage morphology was roughly similar for each species during both periods. That is to say, the number and density of leaves were similar between both images.

Differences observed between the visible spectrums could be related to the level of photosynthetic activity occurring within the leaves during each growing period. A study looking into the spectral absorbance of light relating to photosynthesis found the absorption maxima of chlorophyll a and chlorophyll b (when extracted in diethyl ether) has peaks located at 430 nm and 453 nm respectively (Ustin et al., 2009). This peak chlorophyll absorption corresponds to the band 1 reflectance range from the Pleiades satellite which records wavelengths between 430 – 550 nm. In Figure 16 we can see that much more sunlight is reflected from the leaves in the EOD image than the EOW image in band 1. This indicates that more photosynthetic activity is occurring within the trees during the EOW growing period compared to the EOD.

(30)

Figure 16 - Mean spectral reflectance of each tree species for two different growing periods (a) EOW and (b) EOD

(31)

4.2. INTERNAL SPECIES VARIABILITY

For each species, the highest and lowest mean reflectance values for each band / month combination are presented in Table 8. The highest mean reflectance combining all records within a species was obtained from the Aza species (Max=0.255) and lowest maximum was obtained from the Kha species (Max=0.241). The lowest mean reflectance combining all records within a species was obtained from the Par (Min=0.152) and highest minimum was obtained from the Man species (Min=0.171).

Table 8 – Maximum (MAX) and minimum (MIN) mean (X̅) reflectance values for each species and band. The highest maximum and lowest minimum for each band is marked in green and red respectively.

Species _segments# tree B1 B2 B3 B4 NDVI Average

Aza 33 MAX 0.176 0.152 0.144 0.375 0.430 0.255 MIN 0.083 0.102 0.108 0.232 0.267 0.158 Kha 24 MAX 0.169 0.154 0.146 0.323 0.414 0.241 MIN 0.078 0.098 0.107 0.238 0.267 0.158 Lan 95 MAX 0.189 0.161 0.150 0.346 0.395 0.248 MIN 0.084 0.106 0.110 0.239 0.261 0.160 Man 69 MAX 0.163 0.155 0.143 0.382 0.426 0.254 MIN 0.080 0.103 0.109 0.258 0.305 0.171 Par 108 MAX 0.165 0.147 0.141 0.348 0.417 0.244 MIN 0.079 0.099 0.108 0.228 0.247 0.152 Vit 152 MAX 0.174 0.155 0.146 0.353 0.430 0.251 MIN 0.087 0.107 0.111 0.235 0.263 0.161 Box plots showing the spectral variability within each tree species across each of the spectral bands for the combined model scenario are presented in Figure 17. The variation within each species is largest in the NIR band and NDVI, and lowest in the red band (band 3). There are considerable spectral overlaps between all species in each spectral band. The variation within species is similar for each of the bands, but subtle differences do exist. For example, the variance for Man is higher in the NIR (band 4) compared to the other species. Different sections of wavelengths have been “cut and measured” in sections, known as bands. Each of the bands are different and together form the spectral signature of a feature in question

(32)

Figure 17 - Box-whisker-plots showing the spectral variability within each of the focus species using the combined (2017&2018) training data

(33)

4.3. CHANGE BETWEEN GROWING PERIODS

To visually represent spectral changes of the trees between the EOW and EOD growing periods, two calculations were conducted for each tree segment and graphed as an XY scatter chart. A tree segment is the defined as the uniquely identified area determined to represent the tree i.e. each individual tree. The first calculation was a subtraction of the mean reflectance of the EOW image from the EOD image for each tree segment. Secondly, the standard deviation (sd) from the EOW image was subtracted from the EOD image for each tree segment. The results of this calculation are presented as an x-y scatter graph in Figure 18. The x-axis represents the change in standard deviation for each tree segment between growing periods. The y-axis represents the change in reflectance mean for each tree segment between growing periods.

Figure 18 - Scatter chart showing the variation between the 2018 and 2017 object pixel values (y axis=mean, x axis=standard deviation)

The variation between the reflectance of the EOW and EOD tree segments is roughly similar for

Aza, Kha, Vit and Lan species across all bands. Par and Man appear to have noticeably different

values from the other species in the NIR spectrum (band 4). This is likely to reflect phenological changes within these species occurring between the two image acquisition dates. The results tell us that there is little variety between 4 out of the 6 species between the 2 growing periods, however 2 species spectral signatures change between the growing periods.

(34)

4.4. CLASSIFICATION ACCURACY OF EACH IMAGE COMBINATION

Overall accuracies (OA) and K values of the three scenarios obtained through the RF model iterations and confusion matrix summaries are presented in Figure 19. Classification based on the EOW (OA = 71.10%, K = 0.62) was more accurate in comparison with the EOD (OA = 59.67%, K = 0.47) training data. The combined image (OA = 72.35%, K = 0.64) provided the better classification accuracy than either of the two single images.

Figure 19 - Overall classification accuracies and kappa values from the 2017, 2018 and combined datasets.

(35)

Table 9 - Confusion matrices and statistical metrics for each of the three datasets Correct validation outcomes are marked in green. The highest and lowest producer and users’ accuracy for each dataset is marked in blue and red respectively..

S ce n ar ios Species

Grounded Truthed Data

Total # Producer Accuracy (%) User Accuracy (%) OA (%) K value

Aza Kha Lan Man Par Vit

2018 (E OD ) C lassifie r Re sult s Aza 1 0 13 10 2 7 33 3.03 14.29 59.7 0.47 Kha 0 7 1 6 8 2 24 29.17 43.75 Lan 4 0 51 9 9 22 95 53.68 50.50 Man 1 2 10 26 8 22 69 37.68 41.94 Par 0 6 4 4 86 8 108 79.63 72.88 Vit 1 1 22 7 5 116 152 76.32 65.54 2017 (E OW ) Aza 5 0 13 3 9 3 33 15.15 41.67 71.1 0.62 Kha 0 18 1 0 5 0 24 75.00 90.00 Lan 2 0 58 2 10 23 95 61.05 58.59 Man 3 2 3 41 4 16 69 59.42 74.55 Par 1 0 9 1 95 2 108 87.96 75.40 Vit 1 0 15 8 3 125 152 82.24 73.96 C OMBIN ED Aza 5 0 10 4 9 5 33 15.15 55.56 72.4 0.64 Kha 0 18 1 0 5 0 24 75.00 85.71 Lan 2 0 57 1 11 24 95 60.00 61.96 Man 1 1 2 41 5 19 69 59.42 85.42 Par 1 1 7 1 95 3 108 87.96 74.22 Vit 0 1 15 1 3 132 152 86.84 72.13

Differences between the producer accuracies were observed in all three scenarios ranging from a low of 3.03% (Aza - 2018) to a high of 87.96% (Par - 2017 & COMBINED). The user’s accuracies ranged slightly less from a low of 14.29% (Aza - 2018) to a high of 90% (Kha - 2017).

“Producer accuracy is the probability that a value in a given class was classified correctly.

User accuracy is the probability that a value predicted to be in a certain class really is that class. The probability is based on the fraction of correctly predicted values to the total number of values predicted to be in a class.” (Harris, 2020)

Both Par and Vit performed quite well in terms of producer’s accuracy across all three classification scenarios, with the lowest recorded classification value 76.32% (Vit - 2018). Kha,

Lan and Man were the mid-range performers with producer accuracies in the multi-image scenario

(36)

5 DISCUSSION

5.1. COMPARING IMAGERY

5.1.1. GROWING PERIODS

Image acquisitions across various growing periods can be used to target phenological changes occurring within certain tree species. These changes can include variations to size, brightness and bare ground or shadow within or surrounding the tree crowns. Both the EOW and EOD images were acquired in months with no recorded monthly rainfall (0 mm) (Figure 5) however, the differences in terms of the preceding months’ climatic conditions are quite large. In the study area, May is preceded by the dry season, a period of long, dry and warm weather, whereas October is preceded by the wet season, a period of wetter and cooler conditions. The growing conditions of the preceding months have a large influence on the foliage of trees used in this study (≥ 5 cm DBH), as well as smaller trees and the surrounding understory. There are clear differences in the vegetation components of the two images. In addition, given that the EOD image was acquired in May, it is possible that there have been some phenological changes, such as flower blooming or shedding of leaves, effecting their reflectance brought about by the first seasonal rains. Phenological changes often occur rapidly and for very short periods of time. Therefore they are difficult to capture using this type of satellite sensor that has a set revisit period and depends on cloud coverage (Karlson et al. 2016). This means that there may be some phenological changes that are not accounted for in the selected images.

The difference between the two single image models were most apparent in the visible spectrum (bands 1, 2 and 3). In this spectrum, the mean reflectance values for each species was higher in the EOD than EOW image. This is interesting as it could reflect phenological changes to the trees. However, from previous studies it is known that the NIR and NDVI bands are most positively correlated with foliage structure and not so much with pigment (leaf colour) (Yang et al., 2017). The differences in mean pixel values between images could be related to the external reflectance (i.e. interference) infiltrating the tree crown boundary and effecting the mean. For example, typically in the 2018 image there were a larger number of trees that were considered as “light affected”. This could be due to the strong reflections of bare soil, which is more common in EOD compared to more grassed over areas surrounding the trees in the EOW image. Also, some leaves are dropped in periods of drier weather, this could be a factor adding external interference to the results in EOD.

A previous study comparing Worldview 2 images from October and July in the study area found October to have best overall classification accuracy when using each single image to form a training dataset (Karlson et al., 2016). In this study there was a 10% difference in overall classification accuracy, with only one species having higher classification accuracy in July, Parkia

biglobosa (Table 6). July is a period of peak vegetation productivity in the study area and this was

reflected by the lower spectral variability between species in this month compared to October (Karlson et al., 2016). This compares similarly with this study which October classification accuracy 11.3% higher than those from May.

Combined with findings from past studies in the study area (Karlson et al., 2016), results have assessed tree classification accuracies during three growing periods of the wet season (start (May),