Advanced Deep Learning Architectures for Accurate Detection of Subsurface Tile Drainage Pipes from Remote Sensing Images

2025-04-27 0 0 1.87MB 8 页 10玖币

侵权投诉

Advanced Deep Learning Architectures for Accurate

Detection of Subsurface Tile Drainage Pipes from Remote

Sensing Images

Tom-Lukas Breitkopf∗a, Leonard Hackel∗a, Mahdyar Ravanbakhsha, Anne-Karin Cookeb,

Sandra Willkommenb, Stefan Brodab, and Begüm Demira

aTechnische Universität Berlin, 10623 Berlin, Berlin, Germany

bBundesanstalt für Geowissenschaften und Rohstoﬀe, 13593 Berlin, Berlin, Germany

ABSTRACT

Subsurface tile drainage pipes provide agronomic, economic and environmental beneﬁts. By lowering the water

table of wet soils, they improve the aeration of plant roots and ultimately increase the productivity of farmland.

They do however also provide an entryway of agrochemicals into subsurface water bodies and increase nutrition

loss in soils. For maintenance and infrastructural development, accurate maps of tile drainage pipe locations

and drained agricultural land are needed. However, these maps are often outdated or not present. Diﬀerent

remote sensing (RS) image processing techniques have been applied over the years with varying degrees of

success to overcome these restrictions. Recent developments in deep learning (DL) techniques improve upon the

conventional techniques with machine learning segmentation models. In this study, we introduce two DL-based

models: i) improved U-Net architecture; and ii) Visual Transformer-based encoder-decoder in the framework of

tile drainage pipe detection. Experimental results conﬁrm the eﬀectiveness of both models in terms of detection

accuracy when compared to a basic U-Net architecture. Our code and models are publicly available at https:

//git.tu-berlin.de/rsim/drainage-pipes-detection.

Keywords: Semantic segmentation, tile drainage pipe detection, visual transformer, U-Net, remote sensing.

1. INTRODUCTION

Subsurface tile drainage pipes are an essential agricultural element, maintaining or improving water management

and crop yields of farm land. Depending on location and respective hydro-climatological conditions, signiﬁcant

parts of farm land are equipped with tile drainage systems (e.g. in Denmark ∼50% of the agricultural area [1]).

Multiple installation patterns of subsurface drainage system are in use like herringbone, double main, parallel,

targeted and complex patterns (see [2] for more information).

Besides their positive eﬀects, subsurface drainage pipes also boost – in high pulse-like signals – nutrient

and pesticide loss of soils with short retention time to surface waters. Especially in areas where surface runoﬀ

cannot reach receiving waters and inﬁltrates into depressions, macropore transport to tile drainage pipes was

shown as a dominant pesticide loss pathway of agrochemicals into surface water bodies [3], posing risk to both

human and ecosystem health. In order to improve the understanding of agro-chemical dynamics in soils, to

calculate tile drainage catchment areas and loads for wetland dimension planning, as well as to prepare reliable

eco-hydrological models, it is thus crucial to identify the location of drainage pipes [4]. One typical example for

parallel tile drainage pipes in ﬂat lowland regions is depicted in Fig. 1. The pipes are roughly ten centimeters

in diameter and installed at about one meter below surface [5]. Precipitation (at t0) inﬁltrates rapidly through

the soil above the drainage pipes (t1and t2) due to preferential ﬂow through especially macropores, leading

to relatively low (close to drainage pipes) and relatively high (farther away from the pipes) soil moisture. As

the surface albedo increases with decreasing soil moisture, those diﬀerences in soil moisture become visible. As

shown in Fig. 1, the soil above the drainage pipes appears brighter after a precipitation event, thus producing

distinct features visible at the surface [2,5].

Further author information: (Send correspondence to Leonard Hackel)

Leonard Hackel: E-mail: l.hackel@tu-berlin.de

*These authors contributed equally to this work

arXiv:2210.02071v3 [cs.CV] 1 Nov 2022

Figure 1: Pictorial representation of drainage pipes and the impact on soil color on the surface [5].

The lack of documentation of installed drainage pipes has motivated researchers to explore other detection

methods, such as thermal images and ground-penetrating radar (GPR) [6]. Their practical applicability is

however limited by the restricted data availability. Conventional image processing methods (e.g. edge detection,

decision tree classiﬁcation and image diﬀerencing [7]) have been applied to remote sensing images of bare soils,

obtained after precipitation events [8]. Machine learning methods in particular DL-based methods have recently

shown good performance on drainage pipe detection from remote sensing images [5,9,10]. As an example,

in [5] a DL-based method is introduced using the deep autoencoder-decoder U-Net architecture [11] to detect

drainage pipe from remote sensing images. The U-Net architecture outperforms other remote sensing image-based

approaches using edge detection. The authors suggest that even though the model shows good performance, it

may still be improved by tuning hyperparameters and using a more diverse training set [5].

In this work we aim to improve on the results of the model introduced in [5] and introduce two existing

DL-based models to the problem of drainage pipe detection: i) a modiﬁed U-Net architecture with multiple

reﬁnements (denoted as improved U-Net); and ii) a visual transformer-based encoder-decoder architecture with

skip connections (denoted as TransUNet).

2. METHOD

We adapt two DL-based models to the task of semantic segmentation for tile drainage pipe detection, one based

on an improved U-Net and one based on a Visual Transformer. Both models take as input an image xwith

Cchannels and a H×Wresolution (height and width) and output a grayscale image of the same resolution.

The gray value of a pixel in the output represents the probability of a drainage pipe being in the location of

that pixel. Formally, the goal is to predict a pixel-wise mapping Mwith x7→ M(x)where x∈RH×W×Cand

M(x)∈RH×W×1. The mapping function is trained by minimizing the dice loss function LDice as deﬁned in [5].

The details of each method are presented in the following subsections.

2.1 Improved U-Net

The U-Net architecture [11] is a fully convolutional auto-encoder-decoder (FCN) networks, which contains a

contracting and an expanding path. The contracting path captures context, whereas the expanding path allows

precise localization. Both paths are in a way symmetric to each other. The U-Net can make use of skip connections

that makes it possible for the model to be trained with few images. Detecting tile drainage pipes from RGB

remote sensing images using a deep U-Net architecture was proposed in [5]. Since the U-Net architecture used in

[5] is a basic U-Net, it is not able to utilize multiscale contexts in the latent space at the bottleneck or emphasize

relevant and deemphasize irrelevant information in the residual connections. These issues get addressed in

the improved U-Net architecture. In [12] several U-Net architectures with several adaptions are introduced,

improving the image segmentation performance without signiﬁcant computational overhead. In this work we

investigate one of these architectures and adapt it to the task of drainage pipe detection. This improved U-Net

architecture incorporates several additional modules w.r.t. the basic U-Net. As it is shown in Fig. 2the core

of the architecture is a four-layer U-Net network using 3×3 kernels and a stride of one for all convolutions.

Padding was added, so that input and output dimensions match and batch normalization was added between

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：

AdvancedDeepLearningArchitecturesforAccurateDetectionofSubsurfaceTileDrainagePipesfromRemoteSensingImagesTom-LukasBreitkopfa,LeonardHackela,MahdyarRavanbakhsha,Anne-KarinCookeb,SandraWillkommenb,StefanBrodab,andBegümDemiraaTechnischeUniversitätBerlin,10623Berlin,Berlin,GermanybBundesanstaltfürGeow...

展开>> 收起<<

Advanced Deep Learning Architectures for Accurate Detection of Subsurface Tile Drainage Pipes from Remote Sensing Images.pdf

共8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！

Advanced Deep Learning Architectures for Accurate Detection of Subsurface Tile Drainage Pipes from Remote Sensing 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: