Dealing with Training Data Paucity: One-class versus Binary Classifiers for SAR-based Flood Detection

Abstract

Climate change increases the likelihood of catastrophic flood events, resulting in destruction of cropland and infrastructure, thereby threatening food security and exacerbating epidemics. These dangerous impacts highlight the need for rapid monitoring of inundation, which is necessary to estimate the dimensions of the disaster. An accurate satellite-based flood mapping can support the risk management cycle, from near-real-time rescue and response until post-event analysis. Current remote sensing techniques allow cheap, quick, and accurate flood classifications, using freely accessible satellite-data, for instance, from the Copernicus Sentinel satellites. Indeed, the Synthetic Aperture Radar (SAR) sensor on-board Sentinel-1 (S1), is uniquely suited to flood mapping due to its 24-hour weather independent imaging technology, and is widely used globally due to the open data availability. Binary classifications are widely used to extract flood inundation from SAR data, but due to the large discrepancy in prevalence of flood/non-flood classes in an S1 tile, finding adequate appropriate labelled samples to train classifiers is extremely challenging in addition to being time consuming. Furthermore, the process of training data collection is non-trivial due to a variety of uncertainties in SAR data originating from the underlying land-use and incorrect labeling could lead to gross misclassifications. For example, if the training data does not sufficiently represent the flood surface roughness diversity, large inundated tracts could be missed by the classifier. Consequently, training a binary can be expensive, slow, and compromise on accuracy, since precise labels for both classes are required despite only one class of interest.

One-class classifiers address this issue, by using only samples of the class of interest, i.e. the true positives, making them the perfect choice for flood classification. Even though one-class classifiers have outperformed classical binary classifiers for a variety of use-cases, surprisingly their performance has not yet been benchmarked against the traditional classifiers in flood mapping literature. Accordingly, this study provides the first assessment of one-class classifiers for flood extent delineation from SAR data.

The study area is a coastal part of New South Wales, Australia, where La Niña led to flooding on 12th March 2021. S1 SAR data was used to classify the inundated area using Support Vector Machine (SVM) and Random Forest (RF) for the binary classification and one-class SVM (OCSVM) for the one-class classification. The data inputs and training data for both flood classifications were the same. For validation concurrent cloud-free SkySat optical-data were used.

Preliminary results suggest that one-class classifiers can perform equivalently or better than standard classifiers for flood detection from SAR images given similar volume of training data. Moreover, one-class classifiers offer the advantage of using limited training data and thus result in lower classifier training as well as processing time, without compromising on detection accuracy. In future, global machine learning ready training datasets such as the WorldFloods or Sen1Floods11 dataset can be used to further simplify the training of one-class classifiers . Based on the results obtained in this first benchmarking study, the use of one-class classifiers for flood mapping should be further explored, for a robust performance assessment given different underlying land-uses and geographical regions.

Full text

Fig. 3: Classification results a) –c) provides the results of the whole study area, d) represent the available SkySat ortho scene for validation with validationpoints,
e)-g) enlargement of the blue rectangle in a)-c). Basemap is always the Sentinel-1 image after innundation.
▪Sentinel-1 (VV, Descending) at
10 m for binary flood classification
▪2021/03/12
▪2021/03/24
▪Global Surface Water Mapping
Layer (2020) for masking out
open water bodies with water > 3
months2
▪Limited training data (polygone
low amount of pixels each class)
▪SkySat ortho scene at 0.5 m
(2021/03/26) for 500 stratified
validationpoints
▪SAR-data provides rapid monitoring of floods
▪Collecting training data is a non-trivial due caused by variety of
uncertainties in spatial resolution and references
▪One-class classifiers only requires samples of the investigated class
and can work with limited training data (in context of flood mapping:
Brill et al., 20211)
Dealing with Training Data Paucity: One-class versus Binary
Classifiers for SAR-based Flood Detection
Clara Lößl¹, Tim Landwehr¹, Antara Dasgupta1,2, Björn Waske¹
References:
1Brill, Fabio; Schlaffer, Stefan; Martinis, Sandro; Schröter, Kai; Kreibich, Heidi (2021): Extrapolating Satellite-Based Flood Masks by One-Class Classification—A Test Case in
Houston. In Remote Sensing 13 (11), p. 2042. DOI: 10.3390/rs13112042.
2EC JRC/Google. (n.d.): Global Surface Water. RetrievedApril 16, 2022 from https://global-surface-water.appspot.com
▪F1 score of SVM and RF are the same and the
OCC is lower. RF and SVM have a higher
precision but also a lower recall than OCSVM
▪OCSVM needs just the half of training data than
RF and SVM
▪Recommend use of OCC caused by less time-
consuming preparations and run-time of
classification
Filtered Sentinel-1
Mosaic Binary Flood Classification
with RF, SVM and OCSVM Post-Processing by
Permanent Water Bodies
Evaluation, Hyperparameter
Tuning & Training Validation
¹ Remote Sensing Working Group, Institute of Informatics, University of Osnabrück, Germany (cloessl@uos.de)
² Water Group, Department of Civil Engineering, Monash University, Australia
Workflow
Fig. 1: Performed Workflow for flood detection
Fig. 4: Precision, recall and F1 score of each classifier
a) RF b) SVM c) OCSVM d)
Q: Which Classifier provides the highest performance with limited training data?
Motivation Study Area
Fig. 2: Study area of Windsor, NSW,
Australia (based of Bing Satellite)
Data summary
Conclusion
Lößl, C., Dasgupta, A., Landwehr, T., and Waske, B. (2022). Dealing with Training Data Paucity:
One-class versus Binary Classifiers for SAR-based Flood Detection. Natural Hazards and Earth
System Sciences (Expected submission - July, 2022)
F1 score 95%-
Confidence Interval
SVM 0.86 0.92
RF 0.88 0.92
OCSVM 0.80 0.86
A7.01 Inland Water Storage and Runoff: Modeling, In Situ Data and Remote Sensing
Accuracy Assessment
Tab. 1: Confidence intervals of F1 score
e) f) g)
.
.
.
.
.
.
.