Cross-Geography Generalization of Machine Learning Methods for Classiﬁcation of Flooded Regions in Aerial Images

2025-04-27 0 0 423.4KB 7 页 10玖币

侵权投诉

Cross-Geography Generalization of Machine

Learning Methods for Classiﬁcation of Flooded

Regions in Aerial Images

Sushant Lenka

, Pratyush Kerhalkar

, Pranav Shetty

, Harsh Gupta

, Bhavam Vidyarthi

, and Ujjwal Verma

1, 2

1Department of Electronics and Communication Engg, Manipal Institute of Technology, Manipal Academy of

Higher Education, India

2Department of Electronics and Communication Engg, Manipal Institute of Technology Bengaluru, Manipal

Academy of Higher Education, India

Abstract

Identiﬁcation of regions affected by ﬂoods is a crucial piece of information required

for better planning and management of post-disaster relief and rescue efforts.

Traditionally, remote sensing images are analysed to identify the extent of damage

caused by ﬂooding. The data acquired from sensors onboard earth observation

satellites are analyzed to detect the ﬂooded regions, which can be affected by low

spatial and temporal resolution. However, in recent years, the images acquired from

Unmanned Aerial Vehicles (UAVs) have also been utilized to assess post-disaster

damage. Indeed, a UAV based platform can be rapidly deployed with a customized

ﬂight plan and minimum dependence on the ground infrastructure. This work

proposes two approaches for identifying ﬂooded regions in UAV aerial images. The

ﬁrst approach utilizes texture-based unsupervised segmentation to detect ﬂooded

areas, while the second uses an artiﬁcial neural network on the texture features

to classify images as ﬂooded and non-ﬂooded. Unlike the existing works where

the models are trained and tested on images of the same geographical regions,

this work studies the performance of the proposed model in identifying ﬂooded

regions across geographical regions. An F1-score of 0.89 is obtained using the

proposed segmentation-based approach which is higher than existing classiﬁers.

The robustness of the proposed approach demonstrates that it can be utilized to

identify ﬂooded regions of any region with minimum or no user intervention.

1 Introduction and Application Context

With the increase in the frequency of occurrence of natural disasters, there is a need for better

management and planning of post-disaster relief and rescue efforts. Identiﬁcation of the region

affected by the disaster is one of the crucial information required for planning post-disaster rescue

efforts. The existing method for identifying disaster-affected areas depends on images acquired from

sensors onboard Earth Observation Satellites[10] [7]. However, these images may be affected by

cloud cover and have poor spatial resolution. Besides, the longer satellite visit time may adversely

affect the post-disaster relief and rescue efforts. In recent years, images acquired from Unmanned

Aerial Vehicles have been utilized for various applications such as environment monitoring, damage

assessment etc [12] [11] [1].

For detecting the ﬂooded areas, the aerial images are classiﬁed as ﬂooded or non-ﬂooded depending

on the presence/absence of ﬂooding in the image. In this paradigm, features (Color, textures etc) are

ﬁrst extracted from the images which are then fed to a classiﬁer. In the last decade, several end-to-end

Convolutional Neural Network (CNN) based approaches have been proposed, which eliminates

36th Conference on Neural Information Processing Systems (NeurIPS 2022).

arXiv:2210.01588v1 [cs.CV] 4 Oct 2022

the need for hand-engineering the features [5]. In addition to the classiﬁcation-based approach, a

few studies use the segmentation results to categorize the image as ﬂooded or non-ﬂooded. In this

approach, the images acquired from UAV are segmented and assigned a semantic label (ﬂooded/non-

ﬂooded) to each pixel in the image [2] [3] [6]. For instance, authors in [2] used Fully Convolutional

Networks (FCN) to segment the UAV images and detect the ﬂooded regions. In separate work, in

addition to segmenting the optical UAV images, the digital elevation model was analysed to detect

the ﬂooded region [3]. In another study, an unsupervised segmentation algorithm was utilized to

identify the ﬂooded regions [6]. However, most existing classiﬁcation and segmentation methods

for identifying the ﬂooded regions are limited to a particular region. While [2, 3] analysed images

of ﬂood-prone areas of North Carolina, USA, [6] analysed images acquired in Texas, USA, after

Hurricane Harvey. Extending these methods to new regions such as ﬂood-prone regions of South Asia

would require manually annotating additional images, which is a time-consuming process. There is a

need to develop methods for detecting ﬂooded regions that can be generalized to unseen regions with

minimum manual intervention.

The work proposes two texture features-based approaches to classify UAV aerial images as ﬂooded or

non-ﬂooded across different geographical regions. Indeed, the texture features of pixels corresponding

to water are different from that of other classes (Road, Buildings, greenery, etc), and, therefore can

be used to identify ﬂooded regions. In the ﬁrst approach, an artiﬁcial neural network (ANN) is

proposed in this work which distinguishes the images in texture feature space. In another approach,

a texture feature-based unsupervised segmentation is proposed to ﬁrst segment the images. The

number of pixels corresponding to the ﬂooded region is then used to classify the image as ﬂooded

or non-ﬂooded. The segmentation-based approach is an extension of the method proposed in [6].

A texture-based unsupervised segmentation method was proposed in [6] to assess the severity of

ﬂooding in UAV aerial images of the same geographical region. In this work, we adapt this method to

ensure cross-geography generalization. The proposed method can be utilized to identify the ﬂooded

region and assess the ﬂood severity in a UAV aerial image of any region. In this work, the two

proposed methods are compared with other existing machine learning classiﬁers and segmentation

methods (UNet [9]) to assess the performance of the proposed methods in detecting ﬂooded regions

across different regions.

2 Methodology

This work extends the texture-based unsupervised segmentation approach proposed in [6] for ﬂooded

region identiﬁcation for cross-geography generalization. A summary of this unsupervised segmenta-

tion approach is presented in Section 2.1. In addition, this work proposes a neural network-based

approach to identify ﬂooded regions (Section 2.2 ).

2.1 Texture Based Unsupervised Segmentation

The input image is ﬁrst segmented into different regions using k-means segmentation. The segmen-

tation is based on color (LAB color space) and texture information (Local Binary Pattern (LBP)).

Including texture features ensure a more robust identiﬁcation of regions covered with water. Indeed,

pixels representing water have a distinctive texture appearance compared to other regions (roads,

buildings, etc). The segmented region does not indicate which of the

regions represent the ﬂooded

area. Therefore, each

regions are compared with a reference ﬂooded region using the histogram of

Local Binary Pattern features. Subsequently, the input image is categorized as ﬂooded or non-ﬂooded

based on the number of pixels corresponding to the ﬂooded region. In [6], the reference ﬂooded

the region and the input image were acquired in the same region. However, this work utilizes the

reference ﬂooded image of a completely different region. The use of the unsupervised segmentation

approach along with reference images of a different region ensures that the proposed approach can be

adapted to identify the ﬂooded area of any region with minimum user supervision.

2.2 Artiﬁcial Neural Network

A customized Neural Network is proposed in this work for identifying the ﬂooded regions in UAV

aerial images. The input to the neural network is Local Binary Pattern (LBP) computed on the input

image. Using LBP features instead of gray-scale or color intensities ensures that the network learns

the texture pattern of the input image. The network consists of a sequence of seven dense layers

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Cross-GeographyGeneralizationofMachineLearningMethodsforClassicationofFloodedRegionsinAerialImagesSushantLenka1,PratyushKerhalkar1,PranavShetty1,HarshGupta1,BhavamVidyarthi1,andUjjwalVerma1,21DepartmentofElectronicsandCommunicationEngg,ManipalInstituteofTechnology,ManipalAcademyofHigherEducation,In...

展开>> 收起<<

Cross-Geography Generalization of Machine Learning Methods for Classiﬁcation of Flooded Regions in Aerial Images.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Cross-Geography Generalization of Machine Learning Methods for Classiﬁcation of Flooded Regions in Aerial Images

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: