2. INTRODUCTION
• Global aquaculture accounts for 48% of all fish
production, projected to reach 53% by 2030
• Bangladesh is the fifth largest aquaculture
producer in the world
• Aquaculture output increased 75% between 2002
and 2020, from 1.47 million metric tons to 2.58
million metric tons
3. INTRODUCTION
• Land use competition, food security concerns,
and patchy data create an urgent need for new
aquaculture identification methods
• Publicly available remote sensing products offer
an unprecedented opportunity to quantify food
production at large scales and with a minimal
cost
4. INTRODUCTION
• We evaluated the use of synthetic aperture
radar (SAR) and multispectral data to
detect aquaculture waterbodies in
Southern Bangladesh
6. STUDY AREA
• The proposed framework is
implemented in seven
districts within Southwest and
South-Central Bangladesh
• Area of 17,385 km2
• The dominant land use is
agriculture (62%), followed by
built-up areas (23%),
waterbodies (13%), and
wetlands (2%)
7. DATA
• We collected SAR and multispectral
imagery with 10-m spatial resolution from
Sentinel-1 and 2 missions using Google
Earth Engine (GEE)
• Image collections were obtained from
October 26, 2020, to November 15, 2020
(post-monsoon season)
10. RESULTS
• We evaluated the relative
importance of predictors
• The waterbody area showed
a consistent relevance among
the three classifiers
• CART relied on shape indices
• RF used both shape and
backscatter variables
• SVM used shape, reflectance,
and backscatter information
12. RESULTS
Water polygons from SAR imagery
a) Ground truthing
data
b) Classification Tree
c) Support Vector
Machine
d) Random Forest
e) Unsupervised
classification
f) Ensemble
13. RESULTS
Water polygons from NDWI
a) Ground truthing
data
b) Classification Tree
c) Support Vector
Machine
d) Random Forest
e) Unsupervised
classification
f) Ensemble
15. DISCUSSION
We found that existing approaches for aquaculture
waterbody detection can be improved when
1) generating ensembles for water detection accounting
for individual results from both multispectral and SAR
data;
2) using SAR and multispectral imagery for feature
segmentation purposes; and
3) incorporating backscatter data to train machine
learning classifiers and inform unsupervised methods.
16. DISCUSSION
• Water detection: at least an overall 4% increase
in true positive rates when using the ensemble
water mask compared to the individually
generated masks
• Feature segmentation: limitations within
agricultural and forested environments (very
irregular and interconnected patches or very
large water polygons).
17. DISCUSSION
• SAR data played a critical role in model training and
prediction:
– VV polarization values were used for both water detection and
feature segmentation
– VH polarization values showed a high relative importance in two
machine learning classifiers (i.e., SVM and RF).
• In some cases, the prediction results failed in better
discriminating between aquaculture waterbodies and
other waterbodies.
18. DISCUSSION
Factors Affecting the Performance in Aquaculture
Waterbodies Detection
• Vegetation interference
• Large, irregular polygons resulting from the 10 m
spatial resolution and edge detection approach
prevented the identification of small waterbodies
• Edge detection yielded multiple irregular shapes that
affected the overall aquaculture waterbody detection
19. CONCLUSIONS
• Improved water detection rates, with overall rates
of ~60% and up to ~87% in individual districts.
• C-band SAR-VH information and shape indices
such as eccentricity played important roles in
better differentiating waterbodies.
• Limitations in water detection were mainly
dictated by the dataset’s spatial resolution and
vegetation interference
20. CONCLUSIONS
• Shortcomings in feature segmentation mostly
resulted from poorly defined borders or highly
irregular water polygons
• This affected the transferability of well-performing
supervised classification results into final
predictions (i.e., overall accuracies up to ~79%
for validation to ~24-35% for prediction)