Mapping Olive Plantations Using Sentinel-2 MSI Imagery Case Study: Bashiqa City, Iraq

1
Mapping Olive Plantations Using Sentinel-2 MSI
Imagery Case Study Bashiqa City, Iraq

• Presenter
• Jwan Aldoski
• J Al-Dosky1, H A Mossa2 and F M Hassan2
• 1 Bangor Business School, Bangor University, LL57
  2DG, United Kingdom.
• 2 Physics Department, College of Education,
  Mustansiriyah University, Baghdad, Iraq.

2
Content
3
Introduction

• Agriculture is a vital industry that generates
  profits and employs worldwide.
• Real-time Information for Agriculture is now
  critical key for decision-making.
• Data and Analysis Methods are critical for
  delivering Up-to-date Information.

Olive Plantation Mapping
Information about Olive Plantation is derived
from Statistics, Surveys, and Aerial Imagery .
Recently , Remote Sensing Data is Used
4
Why Remote Sensing Data for Olive Plantation
Mapping ?-The accessibility of medium or
high-resolution satellite data -Free image
acquisitions (some satellite data)-Improve in
visualization, processing, combination methods
and analysis tools
Introduction

• Challenges in Remote Sensing Data (Sentinel-2A
  MSI data) For Olive Plantation Mapping?
• -Large Data
• -Computing Power And Time
• -Cloud Coverage Affects
• - Recorded, Stored, and Data Processed Techniques
• -Spatial and Temporal Resolution
• - Classification Mapping and Evaluation
  Techniques
• - Differentiate it from other Types of Crop

5
Study Goals

• The study goals are
• 1) to assess the capability of Sentinel-2A MSI
  imagery for olive plantation mapping in Bashiqa
  city, and
• 2) to compare the effectiveness of three
  classification techniques (Vector Machine Support
  (SVM), QUEST Decision Tree (QUEST), and K-Means).

6
Martials and Methods

• 1- Study Area
• The study site (Bashiqa city) is in northern
  Iraq, between 3628'12''N and 3625'39.62''N and
  4318'43.98''E and 4322'11.46''E .The climate is
  suitable for olive plantations, wheat corn,
  sunflowers. Olives are now one of the region's
  main cash crops.

7
Martials and Methods
2- Sentinel-2A data

• Sentinel-2A launched in June 2015, and the images
  are freely accessible to the public via
  (https/scihub.copernicus.eu/) 1. Sentinel-2A
  data is defined as 13 spectral bands.
• Sentinel-2A MSI imagery was acquired on April
  12th, 2019, at Level-1C and geocoded (solar
  azimuth 1270, solar elevation 660) using a
  standard TOA reflectance product before being
  pre-processed.

No. Sentinel-2 Bands Name Central Wavelength (µm) Resolution (m) Bandwidth (nm)
1 Coastal aerosol 0.443 60 20
2 Blue 0.490 10 65
3 Green 0.560 10 35
4 Red 0.665 10 30
5 Vegetation Red Edge 0.705 20 15
6 Vegetation Red Edge 0.740 20 15
7 Vegetation Red Edge 0.783 20 20
8 NIR 0.842 10 115
9 Narrow NIR 0.865 20 20
10 Water Vapour 0.945 60 20
11 SWIRCirrus 1.380 60 30
12 SWIR 1.610 20 90
13 SWIR 2.190 20 180
8
Martials and Methods
3- Olive Mapping Workflow
Step One Data Pre-processing
9
Martials and Methods
Google Earth imagery
Step One Data pre-processing

• Landsat 8-OLI at 30 m imagery

Sentinel-2A MSI at 10 m imagery
10
Martials and Methods
3- Olive Mapping Workflow
Step Two Image Classification and Validation
11
Martials and Methods
Step Two Image Classification and Validation
12
Results

• 1-Olive Plantation Map

K-means Technique  
QUEST Technique  
SVM Technique  
Cause of confusion and error

• Classification Technique,
• Remote Sensing Data And Image Acquisition Date
  And Time
• Size Training Sample
• Spectral Features Between Freshly Planted Olive
  And Mature Olive (Ages)

13
2-Matrix Confusion for SVM, QUEST and K-Means
Techniques
Results
Methods LC Classes Ground Truth Samples in Pixels Ground Truth Samples in Pixels Ground Truth Samples in Pixels Ground Truth Samples in Pixels Ground Truth Samples in Pixels Total Classified Pixels UA ()
Methods LC Classes Olive Urban Area Agricultural Barren Grass Area Total Classified Pixels UA ()
K-Means Olive 2176 0 431 32 0 2639 82.46
K-Means Urban Area 0 1611 0 7 170 1788 90.1
K-Means Agricultural 2 0 4815 699 12 5528 87.1
K-Means Barren 1 1 525 3447 134 4108 83.41
K-Means Grass Area 0 42 21 239 3649 3951 92.36
K-Means Total ground truth pixels 2179 1654 5792 4424 3965  
K-Means PA () 99.86 97.4 83.13 77.92 92.03  
K-Means OA () 87.14 87.14 87.14 87.14 87.14 87.14 87.14
K-Means k 0.833 0.833 0.833 0.833 0.833 0.833 0.833
QUEST Olive 2104 0 58 9 0 2171 96.91
QUEST Urban Area 0 1591 0 2 109 1702 93.48
QUEST Agricultural 73 0 5194 747 7 6021 86.26
QUEST Barren 2 1 520 3449 132 4104 84.04
QUEST Grass Area 0 62 20 217 3717 4016 92.55
QUEST Total ground truth pixels 2179 1654 5792 4424 3965
QUEST PA () 95.56 96.19 89.68 77.96 93.75
QUEST OA () 90.13 90.13 90.13 90.13 90.13 90.13 90.13
QUEST k 0.86 0.86 0.86 0.86 0.86 0.86 0.86
SVM Olive 2179 0 184 0 0 2363 92.21
SVM Urban Area 0 1652 0 0 13 1665 99.22
SVM Agricultural 0 0 5317 475 0 5792 91.8
SVM Barren 0 0 291 3949 0 4240 93.14
SVM Grass Area 0 2 0 0 3952 3954 99.95
SVM Total ground truth pixels 2179 1654 5792 4424 3965    
SVM PA () 100 99.88 91.8 89.2 99.67    
SVM OA () 94.63 94.63 94.63 94.63 94.63 94.63 94.63
SVM k 0.93 0.93 0.93 0.93 0.93 0.93 0.93

• OA and K were 87 and 0.83 for K-Means
  classification
• OA and K were 90 and 0.86 94.63 ,0.93 for
  supervised classification approaches (SVM and
  QUEST respectively )  

14
3-Training Sample Sizes
Results

• Once the size of the training dataset became
  large enough (gt 500 pixels/class), both the QUEST
  and SVM methods were equally accurate and
  significantly better than for the K-Means
  technique .
• The SVM and K-Means techniques had higher OA and
  K values once the training sample size was lower.
• When training samples rose, accuracy output
  quickly enhanced the classification techniques of
  K-Means for the QUEST technique.

15
4-McNemar Test Value (Z value) for SVM, QUEST and
K-Means Techniques
Results

• For both the SVM and K-Means methods, the McNemar
  test provided Z values varying from 4.014 to
  7.432 for SVM at higher training data sample
  sizes.
• The values of Z differ depending on the sample
  size of the training sample for both QUEST and
  SVM ,when the sample size rose, the value of Z as
  well increased.
• the K-Means technique did not require a training
  sample as an unsupervised technique

Size of training samples 20 pixels /class 50 pixels /class 100 pixels /class 200 pixels /class 300 pixels /class 400 pixels /class 500 pixels /class
SVM vs. QUEST 17.876 10.825 8.548 6.148 2.864 1.826 - 0.689
SVM vs. K-Means - 4.014 -2.847 1.842 6.019 7.432 7.241 7.087
QUEST vs. K-Means -18.853 -14.21 - 9.215 -9.071 - 4.991 5.279 8.301
16
Conclusions
17

• Thank you
• Any Questions ?