MSc. Thesis (MSCV)

Department of Applied Computer Science University of Bourgogne

Non-destructive assessment of quality and maturity of mangoes based on image analysis

Kumar Ankush

A Thesis submitted in July 2020 for the Degree of MSc in Computer Vision (VIBOT/MSCV)

Master thesis supervised by Dr. Emile Faye and Dr. Julien Sarron of the UPR HortSys of CIRAD.

Abstract

Quality and maturity estimation of mangoes are important for the organization of harvesting date and post-harvest conservation. Although extensive fruit quality estimations exist, they are mostly destructive in nature and available tools for non-destructive estimation are limited. Maturity estimation rely on visual inspection of the fruit, in-hand feel or destructive measure- ment based on quality assessment. Thus, non-destructive tools for an accurate estimation of the quality and the maturity of the fruit have yet to be developed, especially for smallhold- ers. The aim of this study was to develop a tool for non-destructive assessment of quality and maturity of mangoes based on image analysis. This experiment studied 1040 lateral RGB images of 520 mangoes of di↵erent stages of maturity and harvested in two orchards in West Africa. Upon performing digital image segmentation on the images of mangoes, six image fea- tures were calculated with the use of digital image processing functions in MATLAB and four destructive features were taken in consideration. Then, correlations between destructive and non-destructive features of mangoes were explored.

No research is ever quite complete. It is the glory of a good bit of work that it opens the way for something still better, and this repeatedly leads to its own eclipse...

Mervin Gordon Contents

Acknowledgments iv

1 Background 1

1.1 Introduction...... 1

1.2 Problemdefinition ...... 6

1.3 Stateoftheart...... 7

2 Methodology 8

3Results 12

3.1 First part of the objective: segmentation of images ...... 12

3.2 Second part of the objective: classification of images into sunny and shade

side ...... 12

3.3 Third part of the objective: calculation of color and shape image features . . . . 13

4 Discussions 15

A The first appendix 17

Bibliography 23

i List of Figures

1.1 Production of Mangoes, mango steens and guavas in West Africa ...... 3

1.2 Production of Mangoes, mango steens and guavas in Senegal [3] ...... 4

1.3 Production of Mangoes, mango steens and guavas in Cote d’Ivoire [3] ...... 5

2.1 Mango being drilled for pulp collection ...... 10

2.2 Maturity Scale for Mangoes ...... 11

A.1 Mango Sample Batch and segmentation images ...... 18

A.2 Mango Sample Batch and segmentation images ...... 19

ii List of Tables

3.1 Batch accuracy of classification ...... 14

iii Acknowledgments

I take this opportunity to express my sincere and heartfelt gratitude to my Master’s thesis supervisor Dr. Emile Faye for believing in my capabilities and encouraging me to understand and link complex concepts.

I would like to thank Dr. Julien Sarron. I really appreciate your constant support through- out broadening my knowledge and professional development. You taught me to be like a sponge and absorb all the knowledge that comes along the way and that there is no evidence or measure of failure or success if you do not even try.

I am thankful to Dr. Lew and Dr. Fofi for their guidance during the coursework. I am highly indebted to Herma, , Elizabeth and Aurelie for all the administrative and ocial assistance.

Lastly, I express gratitude towards my dear VIBOT batchmates for their unconditional support throughout this Master’s Programme. I will not be able to thank enough my parents who always strenghthen me by teaching to stand tall and fight hard to find the way through the world of uncertainties. I feel blessed to have my wife Pratima Gurung’s immense support and love throughout my journey.

iv Chapter 1

Background

1.1 Introduction

There is a pressing need for fruit maturity estimation for increased food security and less wastage of perishable food in either storage of transportation. According to an estimation of Food and Agriculture Organization of the United Nations (FAO), in 2011 a third or 1.3 billion metric tons of the world’s food was lost or wasted each year [5] with about 88 million metric tons wasted in Europe [6] with associated cost around 148 billion Euros. In order to reduce the food waste proper grading of food is important to decrease food loss and increase profitability of the farmers.

Skin colour is an acceptable and recognized maturity indexing technique for many fruits [20]. Since most fruits show demonstrable changes on the outside with change in their physio-chemical properties, such as hormone secretion and change in polarization of reflected light, change in skin colour is however the most easily perceivable one. Having an index-scale to grade the change becomes useful in plucking, handling and sorting and thus becomes a key step in cultivation management. Producers also strive to prevent products with defective colors from entering the market, as well as to ensure that individual products are packed in batches of similar colour. It is crucial to identify maturity index of fruits since immature fruits exhibit erratic ripening behavior if handled improperly [15]. It almost certainly leads to improper flavor, aroma and texture of the fruit, ultimately leading to its rejection either during transport or in retail.

By addressing these key concerns will impact in three big ways. First, it will reduce food loss. Streamlining the process of harvesting, picking of fruits, appropriate packaging and trans-

1 Chapter 1: Background 2 port will result in less food which is lost before it can be served on a plate. Second, it will reshape the workforce and organize the logistics eciently [16]. Knowing certain packages must be transported on a priority and carefully, will give access to prioritize specific markets and conveyance. Third, it will optimize long term storage and retail. Fruits can be sold at priority in retail depending on their maturity, thus reducing loss to the seller and leading to more value delivered to the consumer.

However, these estimations aren’t easy to perform by humans. In a study by Paulus et al. , even trained operators faced diculties in reproducing the estimation of quality where parameters such as shape, size and colour were involved and termed it as ‘inconsistency’ [26]. As the number of parameters increased, which is an increase in complexity and decision making process, the error in classification also increased.

Area of Computer Vision o↵ers powerful tools for automated inspection of fruits and vegeta- bles. The artificial vision system is capable of achieving beyond human capacity in evaluating long-term process accurately in much shorter time through image capturing devices. It can simplify tedious monitoring process flow that take complex operations to be performed. With use of devices which can ‘see’ electromagnetic waves outside of visible range have brought more data points, thus adding more features for accurate analysis [35]. Hyperspectral analysis can even ‘look into’ food damage through bruise of infection under the peel, helping to isolate such food and hence controlling the spread of disease or release of excess plant hormones from bruised parts of the food.

Choi KH et al. in 1994 in accordance with United States Department of Agriculture pro- posed a classification of tomatoes in 6 classes [7]. Similar attempts were made by Liming X et al. in 2010 for classification of strawberries [17]. Martynenko et al. in 2008 developed a process for estimating changes in density and porosity of ginseng roots during drying, thus eliminat- ing use of scanning electron microscope in imaging [19].The ultimate aim of many computer vision-based inspection systems is to estimate one or more features of the product of interest at a given time and compare them to the consistency that is usually associated with maturity, absence of deformities and blemishes, etc. Other systems are designed to determine the prod- uct’s evolution over time to determine whether a given process is valid.

Mango ( indicaL.) is the world’s 5th most produced fruit (50 metric tonnes) [9], grown mainly in semi-arid, tropical and subtropical areas. The fruit is produced mostly by In- dia (16 Mt) and China (4.3 Mt) in South-East Asia. Mexico (23% of world exports), Brazil (14.3%) and Peru (10.3%) are the main exporting countries [9]. All of Africa accounts for 9% 3 1.1 Introduction while West Africa 4% (1.5 Mt) of the world’s mango production [4]. In West Africa, however, production has doubled over the last decade, marking the country with the highest rate of growth in production in the world.

Figure 1.1: Production of Mangoes, mango steens and guavas in West Africa

Cote d’Ivore and Senegal combined are major mango producing countries in West Africa with 93,267 tonnes and 133,636 tonnes of produce in 2018 respectively [10]. Of these, around 10,000 and 60,000 tonnes were exported to the Europe [4]. A chart representation of the produc- tion of mango, mangosteens and guavas as reported by the FAO has been included as Figure 1.2 and Figure 1.3.

Mangoes are one of the climacteric fruits – fruits that ripen, become softer and sweeter post-harvest. Such fruits are not considered to be of desired eating quality at the time of initial maturity. It takes a ripening period to develop desired taste and texture at the time of con- sumption. The maturing process is driven by genetic, environmental and biochemical events resulting in biochemical and physical changes such as loss of ascorbic acid [14], increase in total soluble solids including sugar and starch [25]; physical changes such as weight, firmness, size and color [24], and aroma changes, nutritional value and flavour.

Hence, estimating right maturity at crucial stage becomes important since the fruits have to be kept in controlled environment to facilitate ripening. Chapter 1: Background 4

Figure 1.2: Production of Mangoes, mango steens and guavas in Senegal [3]

Mango production in developing countries such as West Africa is of main importance for food security. Agriculture is practiced in this area mainly on small farms - about 80 percent of farms in this area are less than 2 Ha [18] and under a variety of perennial-based crop systems. While di↵erent mango-based cropping systems coexist in West Africa, ranging from diversified smallholder orchards to large monospecific commercial orchards [13]. Due to heterogeneities in the tree structure and low management practices, smallholder systems can display high in-field variability in tree yield and fruit quality. For mango trees, fruit development and variation in quality are even more pronounced than for other fruit crops because the species exhibits erratic reproductive activity, i.e. asynchronism and alternate bearing [8]. Reproductive asynchronisms are explained by a spread over time of the phenological stages (flowering and fruiting) and can be observed among trees of the same orchard but also within the same species.

In Senegal, mango represents 60% of fruit production with an annual production of 150,000 metric tonnes harvested from trees grown on surface area of 41,000 hectares. The import po- tential to the European Union was estimated at 225,000 tons, while the local market could absorb half of such production in 2010. Unfortunately, almost 30% of the produce is rejected in the farms itself.

Quality reduction of mangoes occur at various stages namely maturity variation, packaging, 5 1.1 Introduction

Figure 1.3: Production of Mangoes, mango steens and guavas in Cote d’Ivoire [3] transportation and unloading, ripening and storage, and retailing. According to FAO [1] [2], the food loss due to quality reduction because of maturity variation is 15-20% and researchers believe that this is the most crucial point where accurate estimation will lead to better and appropriate packaging, transportation, storage and retailing.

Here, add information on factor that have an impact on the fruit quality and its variability :

Asynchronism (some fruit can be of di↵erent stages and maturity) • Light and microclimate : variability in temperature, humidity and sun exposition [23] • Management practices • Several biochemical changes occur during mango ripening, among which carotenoid biosyn- thesis is one of the most important, as that to express the bicolor change in cv. ‘Manila’. Mesocarp color correlates with rising sugar/acid ratios (TSS/TA) during ripening, while ripen- ing accelerators might modify the fruit softening rates [16]. The relationship between fruit firmness and the sugar-acid ratio (TSS/TA) was proposed for post-harvest ripeness specifica- tion [32]. Fruit maturity can be recognized by distinct attributes which can be destructive or Chapter 1: Background 6 non-destructive.

1.2 Problem definition

The aim of this study was to explore non-destructive tools for quality and maturity estimation performed on low-cost device. Farmers in developing nations have limited to no access to high cost equipments such as NIR image capturing devices, depth and polarimetric cameras. Al- though these image capturing devices capture details which can be used to include more data points for maturity estimation, these equipment almost always have a price index which is out of reach for farmers of developing nations. This is a major challenge in the advancement of computer vision applications in all verticals of market. Technology penetration is di↵erent in di↵erent corners of the world and hence data acquisition and processing on accessible hardware is imperative. Making technology accessible to people from all walks of life is essential for more inclusive development of nation and people. Accessible technology improves quality of life of the people involved in the process and increases trade value between partners. Less expensive however reliable hardware components are also less expensive to maintain and repair, thus re- ducing cost involved in maintenance of the hardware.

Other maturity estimation techniques involve destructive analysis of fruits by pulp and peel analysis, such as total soluble solids, wet and dry weight of peel, fibre content, firmness and texture analysis. These set of analysis are not consistent since fruits even on same trees can mature di↵erently, and hence there’s no protocol on how to select fruits for performing such tests for consistent results. It goes without saying that not all fruits can be destructively ana- lyzed and then be shipped for consumption. Hence, rapid non-destructive estimation capable of being scaled down to be run on low-end hardware will enable quick harvest and movement of packaged goods from remotest corners of the world. Fruit maturity can be recognized by distinct attributes which can be destructive or non-destructive.

Most e↵ective classifiers make use of expensive hardware such as in studies conducted by Naik et al. [4] makes use of thermal imaging, while Zhang et al. [34] made use of infrared lamps. While low cost techiques have been proposed for quality evaluation of tomatoes by Barrios et al. in their exhaustive study [11] and also by Wang et al. [33] for fruit size estimation on mango tree, the image capturing devices haven’t been talked in detail by Barrios et al. and Wang et al. use RGB-depth camera, which is still not low-cost for most. Hence, our study is unique to aim for low cost-components involved in image capturing and quality analysis. 7 1.3 State of the art

1.3 State of the art

Mangoes are harvested from gardens during the summer, and then transported by distribu- tors to di↵erent markets. The distributors need batches of homogeneous quality and maturity according to distance and demand for market price, whereas the intrinsic of such agricultural products di↵er even for one particular variety originating from the same garden at the same time. Due to variability in range, location, and weather conditions at harvest time, the varia- tions become much wider. The grading of mangoes, as mentioned before, is a tedious job and is dicult to maintain constant vigilance. Machine Learning through computer vision promises to objectively perform multiple operations of subject recognition and assessment of characteristics extracted from the subject.

NIR spectroscopy was used for quality analysis of green apples and prediction of ripe-stage mango of eating quality [29].

In this study to classify mango maturity into their sunny and shadow side A Machine Vision Technique for Grading of Harvested Mangoes based on Maturity and Quality [22] Chapter 2

Methodology

For computers, native color space is RGB. However, this color space is device-dependent and not perceptual, meaning that color di↵erences in this color space have no direct relationship with human-perceived color di↵erences. For this purpose, it is more fitting to convert the images from RGB coordinates into other color spaces such as CIELAB, which better matches human vision or HSV, which is better while working with luminescence of the object. Subsequently, texture features extracted from images can provide summary information defined from scene intensity maps which may be related to visual features (coarseness of texture, regularity, homogeneity, distribution of randomness, roughness, among others).

1. 1. Study sites The study was carried out in three orchards from two major production basins in Senegal and Coast, West Africa. We focused on ‘Kent’ cultivar as it is one of the most cultivated cultivar in both study areas. In addition, ‘Kent’ is a major cultivar for export market making the quality and maturity estimation of particular importance for this cul- tivar. Fruiting of mango was continuously observed for the sampled trees. The first orchard (A) was located in the Niayes region (144’ to 150’ N and 166’ to 172’ W), a major agricultural production basin in western Senegal. The Niayes region is part of the Sudano-Sahelian climatic domain and is characterized by an unimodal rainfall from July to September (mean rainfall of 430 mm per year) and a relatively fresh and humid climate generated by the proximity to the Atlantic Ocean (with monthly mean tempera- tures ranging from 16 C to 35 C). In this area, mango trees flower in February and harvest occurs from end –June to beginning of September. Orchard A is of 0.7 ha and belongs to the smallholder diversified cropping system, as described in Sarron et al. (2018) [30]. This orchard is composed of mango trees of various cultivars including export cultivar

8 9

‘Kent’ and ‘’ and more anecdotal local cultivars. Mango trees are grown along with other fruit tree species (citrus, papaya trees, etc.) and occasionally with market gardening underneath the trees. Orchards B and C were located north of the Ivory Coast, in the Poro region (880’ to 1000’ de N and 528’ et 652’ W), the region for mango export. The Poro region belongs to the Sudanian climatic domain also characterized by a rainy and a dry season but with mean rainfall ranging from 1400 mm to 1000 mm (Ko2017). In north Ivory Coast, mango trees flower in January and harvest occurs from mid–April to end of May. Both orchards B and C are monoculture of mango trees of the ‘Kent’ cultivar, covering an area of 10 ha and 7 ha respectively and with an equal tree spacing of 10 m x 10 m.

2. Sampling of mangoes We selected 10, 15, and 15 ‘Kent’ mango trees in orchard A, B, and C respectively in 2019. In each tree, mango fruits have been selected during the fruiting stages in 2019. For trees of orchard A (Senegal), five fruits per tree were selected at 12 di↵erent dates (from Mai 13 to July 28) for a total of 600 fruits sampled. For tees of orchard B and C (Ivory Coast), three to four fruits per tree were selected at five and seven di↵erent dates (from March 30 to April 20) for orchard B and C respectively. Finally, a total of 230 and 290 fruits were sampled in orchard B and C respectively. At each sampling, fruits were randomly chosen on the tree in order to represent the vari- ability of sun exposure conditions. For analysis, mangoes were first imaged on the tree and then harvested for laboratory analysis.

3. Camera characteristics The camera used to capture images was model TG5 manufactured by Olympus Corpora- tion. It is a CMOS based 1/2.33 inch crop-sensor camera with 12 million e↵ective pixels (approximately 12.7 million actual pixels) of 1.54 m pixel size and 2.37 m2 e↵ective pixel area. The camera is capable of 12-bit RAW image capture.

4. Image acquisition Image of each mango fruit were acquired on the tree and after harvest in the laboratory. A first set of images (set 1) was obtained by taking the fruit under external condition directly in the tree. Operation consisted of holding a board with a black colored A4 paper sheet behind fruit on tree (Figure 1). The black background was chosen to facilitate the segmentation of the fruit. A white circle on the backing paper was used as a reference Chapter 2: Methodology 10

for the estimation of the pixel size. In addition, a colorimetric mire was used as a color reference. The camera to fruit distance was maintained at around 50 cm from the fruit. After taking the picture on the tree, the fruit was harvested and carried in the laboratory. There, two set of images, called set 2 and 3, were taken under controlled light and after washing the fruit to remove the dust (Figure). Set 2 corresponded of images of both side of the external fruit. Then, the fruits were cut in half along the major length axis and the internal of both half-fruits were photographed to obtain image set 3. For each laboratory set, the camera was manually held and oriented vertically downwards toward the mango fruits at a distance of 50 cm.

5. Parameter assessment In the laboratory, physical properties of the mangoes were assessed before the fruit was cut. Each fruit was weighted using a precision balance. The fruit weight was assessed using a precision balance. The length and the width of the fruit were measured along the major axis with the application of a numerical vernier caliper (HDCD01200 Ingco, 200 mm 0.01, China). In addition, fruit circumference at major width was assessed with a measuring tape. After the fruit was cut in half (longitudinally along the major length), pulp was sampled as explicated in Figure 2.1. Pulp samples were mixed and the obtained juice was used for chemical assessment.

Figure 2.1: Mango being drilled for pulp collection

Total soluble solids (TSS) in Brix was assessed with a digital refractometer (PAL-1 Atago, precision 0,2 Brix). For fruits from Senegal, the total acidity (TA) and pH of the mango juice were also measured. TA was determined using 10 mL of the pulp juice and by 11

titrating with 0.1 N NaOH to an end point where the colored indicator, phenolphthalein, changed color. TA was reported as percentages of the citric acid equivalent.

6. Maturity class Maturity class of the fruit was determined by expert classification. The expert visually examined the pulp color of the fruit and used the color scale for ‘Kent’ fruit developed by the University of California-Davis and the University of Florida (Figure 2.2). Thus fruit was classified among 5 stages with stage 1 corresponding to an immature mango and stage 5 corresponding to a ready to eat mango.

Figure 2.2: Kent Maturity Scale for Mangoes Chapter 3

Results

3.1 First part of the objective: segmentation of images

We attempted two di↵erent ways of segmentation on the fruit images taken inside the lab. The first one was segmentation using graph cut method in MATLAB’s in-built ‘Image Segmenter’ tool. Although it achieved good segmentation, it required drawing manual ROI around the fruit, which was not possible for 1050 images and centering of images was not uniform across the batches.

We moved on to second way of segmentation using active snakes contour segmentation, again using MATLAB’s ‘Image Segmenter’ tool. This again achieved good segmentation however, the edges of the fruit were not smooth and the shadow under the fruit sometimes caused the contour to not bend close to the fruit. Also, this process again required drawing ROI manually.

The results obtained performing colour thresholding of the fruit images were promising. Since the contrast was good between the background and the fruit, the segmentation achieved was good and retained much information close to the edges of the fruit, something which could not be said about the previous two methods.

3.2 Second part of the objective: classification of mango images into sunny and shade side

Classifying images into sunny and shady was essential to understand which side of the fruit is more appropriate for maturity estimation. Indeed, depending on the light exposure, the

12 13 3.3 Third part of the objective: calculation of color and shape image features fruit will show di↵erent color. On a single fruit, color might vary across the fruit surface, as a side might be expose to the sun whether the other side is in the shadow. However, during the experiment, the two lateral side of the mango have been imaged in the lab and ‘a’ face corresponded to the fruit shoulder on the left ; and ‘b’ face corresponded to the fruit shoulder on the right. Then, a and b faces did not reflect the solar exposure of the fruit. Thus, it was important to estimates which side of the fruit was exposed to the sun and which side was in the shade. For this estimation, we calculated the amount of red pixels per segmented fruit image. Thus, classification of sun-facing, in-shade and insucient information through fuzzy logic classifier resulted in an accurate classification with 96% and above accuracy for batches of images taken between 13 May 2019 to 17 June 2019 (early stages of maturity) with an outlier of batch taken on 24 June 2019 with 89% accuracy. However, the same fuzzy logic classifying ability returns a steady decline in the accuracy from 94% to 54% on mango images taken between 30 June 2019 to 28 July 2019. Batches 1 to 12 had following respective accuracy:

3.3 Third part of the objective: calculation of color and shape image features

After the fruit images were segmented, they were converted into binary image to calculate total length, perimeter and area of the fruit. All of these data was stored in an Microsoft Excel (.xls) file.

The Normalized Di↵erence Index (NDI) was the only colour index computed. This index was described in Payne et al. [27]

G R NDI = G+R where, G and R refer the the green and red bands of the image, respectively. The NDI has been computed on each pixel of the image to obtain a NDI layer of this image (in matrix format). Thus, two summary statistics, the mean (NDI mean) and the standard deviation (NDI sd), have been computed on the NDI layer. The calculated color indices, computed from the NDI, were stored in the same Excel file. We already have morphological and biochemical indices at Senegal which consists of fruit weight, length, width, circumference, Brix, pH of the peel, TSS, TA, and maturity class. Now that segmented images also have numerous indices, we divided those indices further into two groups; one group for running regression with . Chapter 3: Results 14

S. No. Batch date Actual segmented images Classified images Accuracy of classification 1 13 May 2019 100 100 100% 2 19 May 2019 100 96 96% 3 27 May 2019 100 96 96% 4 03 June 2019 100 96 96% 5 10 June 2019 100 98 98% 6 17 June 2019 98 98 100% 7 24 June 2019 98 86 89% 8 30 June 2019 100 94 94% 9 7 July 2019 98 84 87.5% 10 14 July 2019 86 72 83.7% 11 22 July 2019 98 58 59.7% 12 28 July 2019 96 62 64.5%

Table 3.1: Batch accuracy of classification Chapter 4

Discussions

Amidst many successful completion of objectives, there were also challenges faced. We will go through them in sequence.

First, the data collection at the CIRAD lab in Senegal was immaculate with special atten- tion given to physical scale, contrast and colour preservation with a camera which captures less information attributing to it’s smaller CMOS sensor size. The data acquisition can be further improved by fixing camera on a mount vertically above the mounting of the fruit to increase the ‘centering’ in images. Few fruits were inconsistent in framing of the image, which lead to cropping those images one at a time, outside of batch processing. This will reduce time involved in image segmentation and preparing data set for further analysis.

The source of the light can be di↵used as some of the images su↵er from ‘speckle’ phe- nomenon. This makes thresholding and segmenting an image non-uniform and leave ‘holes’ in the segmented image, however, it can be corrected by using simple commands to fill holes.

There was good accuracy achieved in fruit classification in sun-facing and in-shade cate- gories. At the early maturity stages, red pigmentation is elaborate on the surface which faces the sun. Heat from the sun makes the peel develop carotenoids quicker and hence the red/yellow pigmentation on the largely chlorophyll pigments (green) of the peel. We achieved a classifica- tion accuracy of 96% and higher, however the same cannot be said for the same classifier which was ran on batches of mango images of advanced maturity. In the later stages of maturity, much of the fruit peel turns red as carotene synthesis becomes high and hence accurate classification between sun-facing and shade-facing images are low.

15 Chapter 4: Discussions 16

Other improvements can be made in modifying the classifier which takes in account of nor- malized di↵erence index computed from HSV channel 2 of the image as described by Gonzalez and his colleagues [12].

Other authors such as Shamili M (2019) [31] and Riviera V et al. (2013) [28] have demon- strated using CIELAB and HSB colour spaces can be more ecient at classification than RGB. Works of Zhang et al.(2014) [34] and Naik S et al. (2017) [21] can also be taken as an inspiration for detection of defects on the surface of the fruits.

Immediate goals in the study are performing regression using Principal Component Analysis, exploring the classification of mangoes on their maturity on a scale of 1-5, as mentioned earlier. We will also try to explore regression by using Support Vector Machines, Linear Discriminant Analysis / Quadratic Discriminant Analysis. One of the lesser studied supervised classifier is SIMCA or Soft Independent Modeling of Class Analogy, which will be explored for machine determined class estimation. Appendix A

The first appendix

17 Chapter A: The first appendix 18

Figure A.1: Mango Sample Batch and segmentation images 19

Figure A.2: Mango Sample Batch and segmentation images Bibliography

[1] Fao: Case study on the mango value chain in the republic of guyana. Available at = http://www.fao.org/3/I9607EN/i9607en.pdf.

[2] Fao: Case study on the mango value chain in the republic of trinidad and tobago. Available at = http://www.fao.org/3/I9569EN/i9569en.pdf.

[3] Qc. = http://www.fao.org/faostat/en/data/QC/visualize @bookwestafmango, title = CapacityBuildinginDevelopingandEmergingCountries, authour = ElieChrysostome, url = https : //books.google.fr/books?hl = enlr = id = V QSjDwAAQBAJ, journal = CORAF , year = 2018GoogleBooks, 2010.P g.181[Online].

[4] West african mango producers smiling again, 2018. Available at = http://www.coraf.org/2019/07/31/west-african-mango-producers-smiling-again/.

[5] Food loss and food waste, 2019. Available at = http://www.fao.org/food-loss-and-food- waste/en/.

[6] Food waste, 2019. Available at = https://ec.europa.eu/food/safety/food waste en.

[7] K. Choi, G. Lee, Y. J. Han, and J. M. Bunn. Tomato maturity evaluation using color image analysis. Transactions of the American Society of Agricultural Engineers, 38(1):171–176, 1995.

[8] Ana¨elle Dambreville, Pierre-Eric Lauri, Catherine Trottier, Yann Gu´edon, and Fr´ed´eric Normand. Deciphering structural and temporal interplays during the architectural devel- opment of mango trees. Journal of experimental botany, 64(8):2467–2480, 2013.

[9] FAO. FAO. 2019: The State of Food and Agriculture 2019, Moving forward on Food loss and waster reduction. 2019.

20 21 BIBLIOGRAPHY

[10] Food and Agriculture Organization of the United Nations (FAO). Food loss analysis : causes and solutions. 2018.

[11] Abraham Gast´elum-Barrios, Rafael A B´orquez-L´opez, Enrique Rico-Garc´ıa, Manuel Toledano-Ayala, and Genaro M Soto-Zaraz´ua. Tomato quality evaluation with image processing: A review. African Journal of Agricultural Research, 6(14):3333–3339, 2011.

[12] Rafael C Gonzalez, Richard Eugene Woods, and Steven L Eddins. Digital image processing using MATLAB. Pearson Education India, 2004.

[13] Isabelle Grechi, Cheikh Amet Bassirou Sane, Lamine Diame, Hubert De Bon, Aurore Benneveau, Thierry Michels, Virginie Huguenin, Eric Malezieux, Karamoko Diarra, and Jean Yves Rey. Vergers base de manguiers au Senegal : diversity des modles de conception et de gestion. Fruits, 68(6):447–466, 2013.

[14] Yurena Hern´andez, M. Gloria Lobo, and M´onica Gonz´alez. Determination of vitamin C in tropical fruits: A comparative evaluation of methods. Food Chemistry, 96(4):654–664, 2006.

[15] S. N. Jha, A. R.P. Kingsly, and Sangeeta Chopra. Physical and mechanical properties of mango during growth and storage for determination of maturity. Journal of Food Engi- neering, 72(1):73–76, 2006.

[16] Stefanie Kienzle, Pittaya Sruamsiri, Reinhold Carle, Suparat Sirisakulwat, Wolfram Spreer, and Sybille Neidhart. Harvest maturity specification for mango fruit ( L. ’’) in regard to long supply chains. Postharvest Biology and Technology, 61(1):41–55, 2011.

[17] Xu Liming and Zhao Yanchao. Automated strawberry grading system based on image processing. Computers and Electronics in Agriculture, 71(SUPPL. 1):32–39, 2010.

[18] Sarah K. Lowder, Jakob Skoet, and Terri Raney. The Number, Size, and Distribution of Farms, Smallholder Farms, and Family Farms Worldwide. World Development, 87:16–29, 2016.

[19] Alex I. Martynenko. Porosity Evaluation of Ginseng Roots from Real-Time Imaging and Mass Measurements. Food and Bioprocess Technology, 4(3):417–428, 2008.

[20] Yoshitaka Motonaga, Hiroya Kondou, Atsushi Hashimoto, and Takaharu Kameoka. A method of making digital fruit color charts for cultivation management and quality control. Journal of Food, Agriculture & Environment, 2(3&4):160–166, 2004. BIBLIOGRAPHY 22

[21] Sapan Naik and Bankim Patel. Machine Vision based Fruit Classification and Grading - AReview.International Journal of Computer Applications, 170(9):22–34, 2017.

[22] Chandra Sekhar Nandi, Bipan Tudu, and Chiranjib Koley. A machine vision technique for grading of harvested mangoes based on maturity and quality. IEEE Sensors Journal, 16(16):6387–6396, 2016.

[23] Thibault Nordey, Jacques Joas, Fabrice Davrieux, Michel G´enard, and Mathieu L´echaudel. Non-destructive prediction of color and pigment contents in mango peel. Scientia Horti- culturae, 171:37–44, 2014.

[24] Jos´ede Jes´us Ornelas-Paz, Elhadi M. Yahia, and Alfonso A. Gardea. Changes in external and internal color during postharvest ripening of ’Manila’ and ’’ mango fruit and relationship with carotenoid content determined by liquid chromatography-APcI+-time- of-flight mass spectrometry. Postharvest Biology and Technology, 50(2-3):145–152, 2008.

[25] Malkeet S. Padda, Cassandro V.T. do Amarante, Raphael M. Garcia, David C. Slaughter, and Elizabeth J. Mitcham. Methods to analyze physico-chemical changes during mango ripening: A multivariate approach. Postharvest Biology and Technology, 62(3):267–274, 2011.

[26] I Paulus, R De Busscher, and Eddie Schrevens. Use of image analysis to investigate human quality classification of apples. Journal of Agricultural Engineering Research, 68(4):341– 353, 1997.

[27] A. B. Payne, K. B. Walsh, P. P. Subedi, and D. Jarvis. Estimation of mango crop yield using image analysis - Segmentation method. Computers and Electronics in Agriculture, 91:57–64, 2013.

[28] Nayeli V´elez Rivera, Jos´e J Chanona P´erez, Reynold Farrera Rebollo, Jos´e Blasco, Georgina Calder´onDom´ınguez, Mar´ıa de Jes´us Perea Flores, and Israel Arzate V´azquez. Description of maturity stages of mango (mangifera ´ındica l.)‘manila’by image analysis and ripening index.

[29] Sirinnapa Saranwong, Jinda Sornsrivichai, and Sumio Kawano. Prediction of ripe-stage eating quality of mango fruit from its harvest quality measured nondestructively by near infrared spectroscopy. Postharvest Biology and Technology, 31(2):137–145, 2004.

[30] Julien Sarron, Eric´ Mal´ezieux, Cheikh Amet Bassirou San´e, and Emile´ Faye. Mango yield mapping at the orchard scale based on tree structure and land cover assessed by UAV. Remote Sensing, 10(12):1–21, 2018. 23 BIBLIOGRAPHY

[31] Mansoore Shamili. The estimation of mango fruit total soluble solids using image processing technique. Scientia Horticulturae, 249(October 2018):383–389, 2019.

[32] Ana Lucia V´asquez-Caicedo, Sybille Neidhart, and Reinhold Carle. Postharvest ripening behavior of nine Thai mango cultivars and their suitability for industrial applications. Acta Horticulturae, 645:617–625, 2004.

[33] Zhenglin Wang, Kerry B. Walsh, and Brijesh Verma. On-tree mango fruit size estimation using RGB-D images. Sensors (Switzerland), 17(12):1–15, 2017.

[34] Baohua Zhang, Wenqian Huang, Jiangbo Li, Chunjiang Zhao, Shuxiang Fan, Jitao Wu, and Chengliang Liu. Principles, developments and applications of computer vision for external quality inspection of fruits and vegetables: A review. Food Research International, 62:326– 343, 2014.

[35] Manuela Zude, Michael Pflanz, Craig Kaprielian, and Bryan L. Aivazian. NIRS as a tool for precision horticulture in the citrus industry. Biosystems Engineering, 99(3):455–459, 2008.