Aiding Irrigation Census in Developing Countries by Detecting Minor

Irrigation Structures from Satellite Imagery

Chintan Tundia, Pooja Tank and Om Damani

Indian Institute of Technology - Bombay, Mumbai, India

Keywords:

Minor Irrigation Census, Object Detection, Deep Learning, Remote Sensing, Computer Vision.

Abstract:

Minor irrigation structures such as well and farm ponds play very important roles in agriculture growth in

developing countries. Typically, a minor irrigation census is conducted every ﬁve years to take inventory of

these structures. It is essential that an up to date database of these structures be maintained for planning and

policy formulation purposes. In this work, we present the design and implementation of an online system for

the automatic detection of irrigation structures from satellite images. Our system is built using three popular

object detection architectures - YOLO, FasterRCNN and RetinaNet. Our system takes input at multiple res-

olutions and fragments and reassembles the input region to perform object detection. Since currently there

exists no dataset for farm pond and the only publicly available well dataset covers a small geographical region,

we have prepared object detection datasets for farm ponds and wells using Google Maps satellite images. We

compare the performance of a number of state of the art object detection models and ﬁnd that a clear trade-off

exists between the detection accuracy and inference time with the RetinaNet providing a golden mean.

1 INTRODUCTION

In developing countries, irrigation plays an impor-

tant role in farming and agricultural growth (Kirpich

et al., 1999). Minor Irrigation structures account for a

huge part of irrigation infrastructure due to short con-

struction period and low investment required. These

structures, such as wells, check dams, and farm ponds

(NIC , MeitY., 2014), have cultivable command area

up to 2000 hectare, . In India 65% of the agricul-

ture depends on minor irrigation (Frenken, 2012). It

is essential that an up to date database of these struc-

tures be maintained for planning and policy formu-

lation purposes for which minor irrigation census is

conducted every 5 years. The census data is collected

from village level workers/administrators, revenue or

land records and a survey of different government and

private scheme owners. After collection of ﬁeld data,

the data entry is done on an online portal. State gov-

ernments monitor the progress of ﬁeld work, data en-

try and validation work. The validated data is again

examined by the Central Government before the ﬁnal

report generation. Moreover, prior to conducting cen-

sus, training workshops are conducted at central as

well as regional levels. Thus the process of conduct-

ing minor irrigation census involves lot of cost and

effort.

As per (V. K. Bhatia, et. al, 2010) there are sev-

eral obstacles in conducting minor irrigation census.

Census ofﬁcials face difﬁculties like unavailability of

village records, villages being located at remote areas,

and difﬁculty in explaining villagers technical terms.

Sometimes there is a delay in data collection due to

lack of sufﬁcient staff, non-cooperation by farmers,

elections, and ﬂoods. These problems make census

lengthy, error prone and costly. Further, the persistent

fall in groundwater level means that any rapid change

in number of these structures must be detected early

on for the government to take regulatory actions to

prevent competitive extraction and storage of ground

water in farm ponds, leading to the tragedy of the

commons (Prasad and Sohoni, 2018).

This scenario calls for the development of an au-

tomatic system for detection, mapping, and record-

ing of irrigation structures. Using Remote Sensing,

Computer Vision, and Deep Learning techniques, we

have built an online web based object detection sys-

tem that assists users to get locations and counts of

these structures on a GIS based interface. Currently,

our system provides detection of two important struc-

tures, dug wells and farm ponds and we have incor-

porated three deep learning architectures - YOLO,

FasterRCNN and RetinaNet. We next discuss the mo-

tivation behind choosing these structures.

208

Tundia, C., Tank, P. and Damani, O.

Aiding Irrigation Census in Developing Countries by Detecting Minor Irrigation Structures from Satellite Imagery.

DOI: 10.5220/0009421302080215

In Proceedings of the 6th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2020), pages 208-215

ISBN: 978-989-758-425-1

Figure 1: Distinction between a Farm Pond in a Normal

Photograph Vs a Farm Pond in a Satellite Image.

Figure 2: Farm Ponds Can Be Confused with a Farm or an

Industrial Unit or a Stone Quarry.

Farm Ponds: It is a surface structure dug out

in the earth, usually square or rectangular in shape,

which harvests rainwater for future use as a drought

prooﬁng tool. Construction of farm ponds have been

receiving a great push from the Indian governmentt

for the past few years (Abhay N, 2016), but serious

concerns are also being raised against this drive. The

use of farm ponds has long drifted from its objective

of storing rainwater for protective irrigation. In sev-

eral areas, farm ponds are being used as storage tanks

for pumped out groundwater exposing this precious

underground resource to evaporation losses (Prasad

and Sohoni, 2018). Thus accurate and timely detec-

tion and geo-tagging of farm ponds is essential.

Automatic detection of farmponds is a challenging

task. Any given time, some farm ponds are dry being

empty while others are wet being ﬁlled with water.

These may or may not have plastic lining. Moreover,

several objects can be confused with farm ponds such

as farms, industrial units, and stone quarries as seen

in Figure 2.

Dug Wells: It is a groundwater extraction struc-

ture created by digging or drilling to tap the underly-

ing aquifer. It is an important source of potable water

in rural areas where they are used for multiple pur-

poses. Also, in developing countries, the estimation

of groundwater level in rural area is done with the

help of observation wells.

Detecting dug wells from satellite images is also a

challenging problem. These are very small structures

contained in an area of approximately 40 x 40 pixels

in a satellite image of spatial resolution 15 - 30 cm.

The wells in satellite images look like dark circular

shaped objects with a thin white lining around it. An-

other challenge is that generally they look very simi-

lar to trees or shadows of the trees. The most difﬁcult

ones are those which are found next to or covered by

the shadow of a tree.

2 RELATED WORK

Object detection refers to the task of localizing all rel-

evant objects present in an image (Russakovsky et al.,

2015). Detecting objects from satellite images is a

challenging task and is receiving signiﬁcant attention

in recent years. It suffers from several difﬁculties in-

cluding the large variations in the visual appearance

of objects caused by viewpoint variation, occlusion,

background clutter, illumination, shadow, etc. (Cheng

and Han, 2016). Due to low resolution in satellite

images, earlier it was difﬁcult to detect separate land

use and land cover objects. With advances in remote

sensing, satellites like IKONOS, SPOT-5, and Quick-

bird have been providing very high resolution (VHR)

aerial images with detailed spatial and textural infor-

mation. Due to progress of deep learning, object de-

tection in satellite imagery has found several applica-

tions such as building detection ((Vakalopoulou et al.,

2015),(Yuan, 2018)), road detection (Saito and Aoki,

2015), solar panel detection((Yuan et al., 2016)), ve-

hicle detection ((Dorrer et al., 2019),(Audebert et al.,

2017)). Most of these works use traditional convo-

lutional neural networks and require large dataset to

train the models.

In recent years, object detection techniques have

greatly improved in terms of accuracy and speed

due to better image representation structures and im-

proved deep learning architectures. Latest deep learn-

ing based detectors can be categorized into one-stage

and two-stage detectors. In two-stage detectors, the

ﬁrst stage produces a sparse set of candidate object

proposals, and the second stage classiﬁes object pro-

posals into target classes or background. Selective

Search(Uijlings et al., 2013), Spatial Pyramid Pool-

ing Network(SPPNet) (He et al., 2014), Feature Pyra-

mid Network (FPN) (Lin et al., 2017) and Region

based Convolutional Neural Network(RCNN) (Gir-

shick et al., 2014) are the most prominent two-stage

detectors. RCNN became particulary popular for the

region based tasks and FastRCNN (Girshick, 2015)

and FasterRCNN (Ren et al., 2017) are evolved ver-

sion of RCNN with improved detection speed. While

two-stage detectors dominate the modern object de-

tection they lack in speed compared to the one-stage

detectors. In one-stage detectors, the division of input

image into regions and probabilistic prediction of ob-

ject presence with candidate bounding box is done si-

multaneously. These detectors consist of models like

Single Shot Multibox Detector (Yang et al., 2019) and

You Only Look Once (YOLO)(Redmon et al., 2016).

This improvement in detection speed comes at the

expense of localization accuracy compared to two-

stage detectors, especially for small objects. How-

Aiding Irrigation Census in Developing Countries by Detecting Minor Irrigation Structures from Satellite Imagery

209

ever, there is a recent one-stage detector RetinaNet

(Lin et al., 2017) which is able to match the speed

of previous one-stage detectors while surpassing the

accuracy of all existing two-stage detectors.

In this paper, we compare and evaluate three state-

of-the-art detectors FasterRCNN, YOLO and Reti-

nanet on well and farm pond datasets developed by

us. These methods cover the spectrum of accuracy vs

detection speed trade-off.

3 SYSTEM ARCHITECTURE

Our system is a web-based application which is in-

spired by and a signiﬁcant improvement over the

SIAS - Satellite Image Annotation System (Wagh

et al., 2019). SIAS supported the well detection from

satellite images in a ﬁxed-sized (640x640 pixels) win-

dow and performed detection in small regions only

whereas our system performs detections over large ar-

eas of size of cluster of many villages. Most impor-

tantly, compared to SIAS system which takes roughly

5 minutes to download and process 100 images, our

multi-threaded architecture takes only 10 seconds. In

addition, our system provides a better user interface

thus improving user experience.

The architecture of our system is shown in Fig-

ure 5. The system has features to search and select

an area from the map interface. The users can also

provide a shape-ﬁle of the concerned region as input.

Detection is performed by loading model’s weights

into the architecture from the model repository which

contains trained weight ﬁles of detection models and

results are obtained by passing images of the target

area to the model. The detection result is provided

as the geo-referenced bounding-boxes i.e. the pixel

coordinates of the object are converted to latitude-

longitude points. The geo-referenced bounding boxes

of the detected objects are then overlaid on the map

interface and thus visible to the users on the map it-

self. The system has a feature to take feedback from

the user about the correctness of the detection. User

can also annotate the missing objects. Moreover, the

users have an option to download the report contain-

ing latitude-longitude coordinates of the detected and

annotated objects in the form of a comma-separated

values(CSV) ﬁle. The system’s backend is built us-

ing Django Framework and frontend is built using

HTML, JavaScript and CSS.

3.1 Client Architecture

As shown in Figure 5, the client side module provides

a map interface. The user can also specify area in

Figure 3: System Architecture of the Proposed System.

the form of shapeﬁle. The user can select an object

type and a detection model. All the detection models

discussed later in section 4.2 are available for users

to choose from. The details of modules involved in

client architecture are as listed below:

Search Area: It provides users an online platform to

select any area for detection. It creates a ﬁxed-sized

window which contains the satellite view of any

particular area. Users can search for the required area

by dragging the map on the window or by typing

the name of the area in the search box provided.

Output by the detection model is overlaid on the

map interface window in the form of a marker and

bounding box.

Batch Submission - Shapeﬁle: The system also

supports shapeﬁle input. Shapeﬁle is a ﬁle that

contains the boundary of the village or boundary of

the area that needs to be selected for detection in the

form of vector data.

Feedback and Annotation: The System gives a fa-

cility to add missing objects. User can draw a bound-

ing box on the map interface or can put a marker by

clicking on the map interface. The system also sup-

ports votes from the user. Additionally the user can

also remove any misidentiﬁed objects.

3.2 Server Architecture

Server architecture consists of the following modules

that handles requests from user and perform detection

on the inputs given from client side:

Tiling Module: It accepts the input area provided by

users which can be an area selected in the map in-

terface window or provided by a shape-ﬁle. As our

model works on satellite view image with a resolu-

tion of 0.3m/pixel, we divide the required area into

tiles of size 640x640 pixels and download the satel-

lite images at that resolution. To get the center points

GISTAM 2020 - 6th International Conference on Geographical Information Systems Theory, Applications and Management

210

Figure 4: Process to Get Tiles from a Shapeﬁle.

Figure 5: Screenshot of Proposed System.

of these tiles, we follow the following steps:

1. Take the latitude-longitude coordinates of the

bottom-left corner and top-right corner of the area.

2. Calculate latitude and longitude offset based on the

resolution required for each tile and the size of the

tile in terms of pixels.

3. Starting with the point in the bottom-left corner

add offset to it to get the next point. Continue

this process until we reach the top-right corner and

cover the entire area as depicted in the Figure 4.

We then download images centered around the points

obtained as seen in the Figure 4. These images are

then provided to object detection module for further

process of detection. We observed that one of the

major bottlenecks in SIAS was the time required to

download satellite images. This becomes even worse

when we download a large number of images. De-

pending on the size of the village it requires ∼ 600

to 2000 images to cover the entire area. SIAS took

∼ 5 minutes to download and detect objects for 100

satellite images. Furthermore, we observed that while

downloading images, CPU remains idle.

We use multithreading for downloading and

detection tasks. Our problem is similar to the

famous producer-consumer problem. In our case

the producer is the process to download images

and consumer is the process to perform detection

on each downloaded image. This reduces the time

for downloading and detection in images from 301

seconds to just 22 seconds for 100 images. Our

system uses 8 threads to download the images and 8

threads to process and predict object from images.

Object Detection Module: Object detection module

takes object type to detect ( i.e. Well, Farm Pond

etc.), and detection model from user and image tiles

provided by tiling module. It uses trained model

repository to get pre-loaded weights and network

architecture of the required object detection model. It

processes the images and gets the pixel coordinates

of the detected object and bounding box for the same.

These coordinates are converted into geo-referenced

points and then output is provided to client view.

User Feedback and Annotation Module: The given

feedback by the user about the correctness of the

detection data is stored in the database as human-

veriﬁed objects. Also manually added objects which

were missing in the detection results, are stored in the

database. This data can be used for further training of

the model to improve the accuracy.

Aiding Irrigation Census in Developing Countries by Detecting Minor Irrigation Structures from Satellite Imagery

211

4 OBJECT DETECTION

EXPERIMENTS

4.1 Dataset Description

There is a scarcity of public dataset of minor irrigation

structures. We introduce two datasets with respect to

the two agricultural structures, i.e. Dug-Wells Dataset

and Farm-Ponds Dataset created using data collected

from all the districts with agriculture of the state.

Creating a dataset using satellite images is not

straightforward. It involves choosing set of the im-

ages of target area, cropping the images in required

size and annotating objects in the images. To ease

this process, we created a Satellite Image Labeling

tool that helps an annotator to perform all the above

tasks efﬁciently. This tool provides an interface to

navigate through a map layer, annotate objects using

different shapes like circle and polygon on the layer

and then save images and annotations to the database.

The interface has been built using Django as back-

end and HTML, Javascript, JQuery and Google Maps

APIs(Google Static Maps, 2019) for frontend.

We used the geo-tagged data of works done by

the Jalyukta Shivar Abhiyan (MRSAC, 2015). which

is the ﬂagship programme of the state of Maharash-

tra government of India, aim of which is to make

5,000 villages free of water scarcity by installing or

strengthening decentralized water bodies like wells

and farm ponds to enhance the groundwater recharge

(Anjali Marar, 2019). The labeling interface was cus-

tomized to set map layer according to geo-tagged lo-

cations thereby reducing the search space required by

annotators. The annotations were done by many dif-

ferent experts to prevent bias in the data. The details

of the datasets are as follows:-

• Dug Wells: Dug Wells being hollow circular ob-

jects are annotated using shape as circle. The

dataset consists of total 1011 images consisting of

1614 wells from 34 districts of the state. Each im-

age size is of 640x640 pixels and is extracted from

Google Static Maps taken at a zoom level of 19.

This dataset can be used for image classiﬁcation,

object detection and instance segmentation tasks.

• Farm Ponds: Farm Ponds are found to be in dif-

ferent sizes, shapes and types. The annotations

are done using polygons. Since both dry and wet

farm ponds can be lined or unlined we get four

classes of objects:

– Dry Farm Pond - Lined and Unlined

– Wet Farm Pond - Lined and Unlined

The dataset consists of 1018 images consisting

of 370 instances of each type totaling 1480 farm

ponds from 34 districts of Maharashtra. Each im-

age size is of 640x640 pixels and extracted from

Google Static Maps taken at a zoom level of 18.

Apart from our purpose of detection this dataset

can also be used for multi-class image classiﬁca-

tion and multi class instance segmentation tasks

4.2 Model Architectures

FasterRCNN: Faster RCNN(Ren et al., 2017) is an

evolved model from the Region-based CNN (RCNN)

family of detectors and it is the modiﬁed version of

Fast RCNN (Girshick, 2015) with the main difference

being the use of CNN based Region Proposal Net-

work(RPN) instead of Selective Search to generate

a set of region proposals. Faster RCNN model

architecture consists of three main parts: Convolution

layers that are responsible for extracting features,

RPN for region proposals and fully connected layers

for classifying objects and predicting the bounding

boxes(regression).

We implement Faster RCNN with ResNet-50 back-

bone having 50 convolutional layers with skip

connections and adapt stochastic gradient descent

optimizer with an initial learning rate of 0.005, the

momentum of 0.9 and weight decay rate of 0.0005.

We train the model for 10 epochs of batch size 3.

YOLOv3: You Only Look Once(YOLO) (Redmon

and Farhadi, 2018) is one of the fastest object detec-

tion model and belongs to the class of single-stage

models. Yolov3 is the latest model in the YOLO fam-

ily known for its fast real-time multi-object detection.

The main improvement in YOLOv3 is that it uses

three different scales to detect objects of different

sizes. It adapts a FPN(Feature Pyramid Network)-

like structure for multi-scale detection. The use

of residual blocks (ResNet-like structure) in this

model simpliﬁes the complexity of feature learning.

Tiny YOLOv3 is a simpliﬁed version of YOLOv3

which reduces the depth of the convolutional layer

and therefore is lighter and suitable for real-time

detection at the cost of reduced accuracy.

We follow the implementation of (Redmon and

Farhadi, 2018) for YOLOv3 and Tiny YOLOv3. We

do not rescale the input and use the original input size

of 640x640. We use learning rate of 0.001 and Adam

optimizer and train the model for 100 epochs with a

batch size of 4 for both the models.

RetinaNet: RetinaNet(Lin et al., 2017) is another

single-stage object detection model with FPN(Lin

et al., 2017) and Focal loss for dense object detection.

The FPN above the ResNet network helps to generate

GISTAM 2020 - 6th International Conference on Geographical Information Systems Theory, Applications and Management

212

a multi-scale feature pyramid, two subnetworks above

this backbone is used for classiﬁcation and bounding

box regression. The use of Focal loss helps to focus

more on difﬁcult and mis-classiﬁed samples and to re-

duce the relative loss on well-classiﬁed samples.

We follow the implementation of (Lin et al., 2017).

We experiment this model with ResNet-101 and

ResNet-50 network architecture. We train our model

for 100 epochs and batch size of 16 with learning rate

1e-5 and Adam optimizer. we use the input image size

of 640x640 for both the models.

4.3 Implementation Details

For training the models, we use transfer learning ap-

proach where we adapt pre-trained models (as dis-

cussed in subsection 4.2) and ﬁne-tune using our well

and farm pond datasets. Both the datasets were split

into randomly choosing 80% train and 20% test sam-

ples i.e. 809 training and 202 test examples for Wells

dataset and 814 training and 204 test examples. Back-

bone of all the architectures are loaded by pre-trained

weights on MS-COCO dataset(Lin et al., 2014). All

the models were trained on Tesla K80 GPU with

12GB VRAM.

5 RESULTS

5.1 Evaluation Metric

For evaluation and comparison of the models experi-

mented on our dataset, we use Pascal-VOC 2012 met-

ric of object detection (Everingham et al., 2012). In-

tersection Over Union (IOU) is a measure that eval-

uates the overlap between two bounding boxes. It

requires a ground truth bounding box B

and a pre-

dicted bounding box B

. With the help of IOU we can

tell if a detection is acceptable (True Positive) or not

(False Positive). IOU is given by the overlapping area

between the predicted bounding box and the ground

truth bounding box divided by the area common be-

tween them:

IOU =

(area(B

∩ B

)

(area(B

∪ B

)

• True Positive (TP): IOU >= threshold

• False Positive (FP): IOU < threshold

• False Negative (FN): A ground truth not detected

• True Negative (TN): Correct detection of a nega-

tive instance. Concept inapplicable in our setting

Precision is a measure of the ability of a model to

identify only the relevant objects. It is the percentage

of correct object detection among all detection. Recall

is the ability of a model to detect all the relevant ob-

jects (all ground truth bounding boxes). It is the per-

centage of the relevant objects that are correctly de-

tected. To understand the performance of object de-

tectors we use the metric the Area Under the Curve

(AUC) of the Precision x Recall Curve (Everingham

et al., 2012). Average Precision (AP) is obtained

by interpolating the precision at each recall level r,

taking the maximum precision whose recall value is

greater or equal than r + 1. This estimated area under

the curve is calculated for each object class. Finally

the mean of AP of all classes gives mean Average Pre-

cision (mAP), the metric being used to compare dif-

ferent object detectors.

5.2 Test Evaluation

We perform evaluation on test examples as per the

split as deﬁned in section 4.3) and compare archi-

tectures deﬁned in section 4.2 . For comparison be-

tween the model architectures we report their infer-

ence speed in terms of GPU time and accuracy in

terms of AP as deﬁned in section 5.1. We have set

IOU threshold to 0.5 to deﬁne true positives for all

the architectures. All models provide a conﬁdence

score for each predicted bounding box which tells

how conﬁdent the model is about the bounding box.

We tested different conﬁdence scores from 0.001 to

0.5 and found that 0.5 gives least number of false pos-

itives. All test results have been reported on conﬁ-

dence score 0.5.

Table 1: Results on Well Dataset. FasterRCNN Gives

the Best Accuracy but Performs Poor in Terms of Speed

Whereas YOLO Models Though Being Faster Lags in Ac-

curacy.

Model AP

Inference

time

(ms/image)

YOLOv3 61.0 % 60.5

tinyYOLOv3 82.8 % 18.9

FasterRCNN 94.0 % 146.6

RetinaNet (Resnet50) 86.9 % 64.1

RetinaNet (Resnet101) 89.5 % 84.4

Table 1 and Table 2 show overall performance of all

the models on Well dataset and Farm Ponds dataset

respectively in terms of accuracy and speed. In terms

of accuracy, FasterRCNN performs the best at 94%

for wells and 95.9% accuracy for farm ponds whereas

YOLOv3 has the lowest accuracy in both the datasets.

Figure 6 and Figure 7 show the detection results ob-

tained on all models on well dataset and farm pond

dataset respectively. FasterRCNN result has the cor-

rect prediction with tighter bounds with ground truth

Aiding Irrigation Census in Developing Countries by Detecting Minor Irrigation Structures from Satellite Imagery

213

Figure 6: Results on Well Dataset. In Left to Right Order: YOLOv3, tinyYOLOv3, RetinaNet(ResNet50), Reti-

naNet(Resnet101), FasterRCNN. Green Rectangles Denotes Ground Truth and Red Rectangles Denotes Prediction. In

YOLOv3 and tinyYOLOv3 Some Wells Have Been Missed out.

Figure 7: Results on Farm Pond Dataset. In Left to Right Order: YOLOv3, tinyYOLOv3, RetinaNet(ResNet50), Reti-

naNet(Resnet101), FasterRCNN. Green Rectangles Denotes Ground Truth and Red Rectangles Denotes Prediction.

whereas other models either have False negatives or

False positives. Note that our results do not conﬁrm

with (Lin et al., 2017) where they had achieved best

accuracy on RetinaNet models trained compared to

FasterRCNN on COCO dataset whereas in our case

FasterRCNN still leads in accuracy. RetinaNet mod-

els lag by a huge margin for wells but show competi-

tive accuracy for farm ponds although we found some

results on farm ponds with spurious detection. As

seen in Figure 8 farm and bigger sized wells were de-

tected as farm ponds. RetinaNet(Resnet101) though

having more layers does not necessarily perform bet-

ter than RetinaNet(Resnet50) atleast in case of farm

ponds. Also YOLOv3 misses out many wells and

gives poor recall value thus affects the average pre-

cision. It seems that farm ponds achieve good accu-

racy compared to wells for all the models. This may

be because wells are smaller in size as compared to

farm ponds and more challenging to predict due to the

presence of similar looking structures like trees and

shadow of trees which accounts to more false posi-

tives.

In terms of speed, as seen in the table, YOLO

models are faster in inference time as compared to

FasterRCNN and RetinaNet models though strug-

gling with accuracy. We can see that tinyYOLO

model takes the least amount of time in both the

datasets due to tinyYOLO’s lighter design and less

number of convolutional layers. Though FasterRCNN

performs very well in case of accuracy but lags in

inference time due to its structure of two stage de-

tection. RetinaNet known for ﬁner accuracy and

speed has inference time similar to YOLO mod-

els for RetinaNet(ResNet50). RetinaNet(Resnet101)

has decent accuracy but is slower compared to Reti-

naNet(Resnet50). Since, RetinaNet(Resnet50) has

accuracy simil ar to FasterRCNN and speed similar to

YOLOv3, it provides a sweet spot between accuracy

vs speed tradeoff and can be considered as a choice

when both speed and accuracy are important.

Table 2: Results on Farm Pond Dataset. FasterRCNN Has

the Slowest Inference Time among All the Models Whereas

tinyYOLOv3 Is the Fastest.

Model AP

Inference

time

(ms/image)

YOLOv3 81.7 % 63.0

tinyYOLOv3 92.2 % 18.4

FasterRCNN 95.9 % 151.2

RetinaNet (Resnet50) 95.2 % 62.7

RetinaNet (Resnet101) 94.3 % 99.7

Figure 8: Some Spurious Detections Leading to False Pos-

itives in Results of RetinaNet(Resnet50) on Farm Pond

Dataset. Green Rectangles Denote Ground Truth, Red De-

note Prediction.

GISTAM 2020 - 6th International Conference on Geographical Information Systems Theory, Applications and Management

214

6 CONCLUSIONS

We have built a web-based innovative system for au-

tomatic recognition of structures like wells and farm

ponds. Our system has better accuracy as well as

faster detection time compared to earlier systems. We

have introduced two new public datasets, using which

we explore the trade-off between object detection ac-

curacy and inference time. We ﬁnd FasterRCNN to

be giving the best accuracy though very high infer-

ence time, tinyYOLOv3 to be the fastest but lagging

in accuracy, and RetinaNet providing the golden mean

having both good accuracy as well as reasonable in-

ference time.

REFERENCES

Abhay N (2016). How to get shettale subsidy for farm pond

in 2016. URL: http://indiamicroﬁnance.com/shettale-

subsidy-farm-pond-2016.htm. Online; Accessed 15-

02-2019.

Anjali Marar, T. I. E. (2019). Explained: Jalyukta shivar key

for maharashtra, but still has a long road ahead. URL:

https://indianexpress.com/article/explained/jalyukta-

shivar-key-for-maharashtra-but-still-has-a-long-road-

ahead/. Online; Accessed: 2019-12-20.

Audebert, N., Le Saux, B., and Lef

evre, S. (2017).

Segment-before-detect: Vehicle detection and classi-

ﬁcation through semantic segmentation of aerial im-

ages. Remote Sensing, 9(4).

Cheng, G. and Han, J. (2016). A survey on object detection

in optical remote sensing images. ISPRS Journal of

Photogrammetry and Remote Sensing, 117.

Dorrer, G., Koriukin, M., Yushkova, S., and Sviridova, L.

(2019). Vehicle detection in aerial images. IOP Conf.

Series: Earth and Environmental Science.

Everingham, M., Van Gool, L., Williams, C. K. I., Winn,

J., and Zisserman, A. (2012). The PASCAL Visual

Object Classes Challenge 2012 Results.

Frenken, K. (2012). Irrigation in Southern and Eastern Asia

in ﬁgures - AQUASTAT Survey 2011.

Girshick, R. (2015). Fast r-cnn. In 2015 IEEE Intl. Conf.

on Computer Vision.

Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2014).

Rich feature hierarchies for accurate object detection

and semantic segmentation. In IEEE Conf. on Com-

puter Vision and Pattern Recognition.

Google Static Maps (2019). Google maps javascript api

v3 reference. URL: https://developers.google.com/

maps/documentation/ javascript/reference/3.37. On-

line;Accessed 04-January-2020.

He, K., Zhang, X., Ren, S., and Sun, J. (2014). Spatial pyra-

mid pooling in deep convolutional networks for visual

recognition. Lecture Notes in Computer Science.

Kirpich, P. Z., Haman, D. Z., and Styles, S. W. (1999). Prob-

lems of irrigation in developing countries. Journal of

Irrigation and Drainage Engineering, 125(1).

Lin, T., Doll

ar, P., Girshick, R., He, K., Hariharan, B., and

Belongie, S. (2017). Feature pyramid networks for

object detection. In IEEE Conf. on Computer Vision

and Pattern Recognition.

Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Dollar, P.

(2017). Focal loss for dense object detection. In The

IEEE Intl. Conf. on Computer Vision.

Lin, T.-Y., Maire, M., Belongie, S., Bourdev, L., Girshick,

R., Hays, J., Perona, P., Ramanan, D., Zitnick, C. L.,

and Doll

ar, P. (2014). Microsoft coco: Common ob-

jects in context.

MRSAC (2015). Maharashtra remote sensing ap-

plication centre. jalyukt shivar , water con-

servation department, mantralaya, mumbai.

URL:http://mrsac.maharashtra.gov.in/jalyukt/. On-

line; Accessed: 2019-09-15.

NIC , MeitY. (2014). National informatics centre. ﬁfth

census of minor irrigation schemes. URL: http://

mi5census-data.nic.in/nologin/introduction.action.

Online; Accessed: 2019-09-20.

Prasad, P. and Sohoni, M. (2018). Farmponds for horticul-

ture: Boon or curse? analysing impact on farm prof-

itability, resource sustainability and social welfare. In

Proc., 36th Intl. Conf. of the System Dynamics Society.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time ob-

ject detection. In Conf. on Computer Vision & Pattern

Recognition.

Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental

improvement. arXiv.

Ren, S., He, K., Girshick, R., and Sun, J. (2017). Faster

r-cnn: Towards real-time object detection with region

proposal networks. volume 39.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,

Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-

stein, M., Berg, A. C., and Fei-Fei, L. (2015). Im-

agenet large scale visual recognition challenge. Intl.

Journal of Computer Vision (IJCV), 115(3).

Saito, S. and Aoki, Y. (2015). Building and road detection

from large aerial imagery. In Lam, E. Y. and Niel,

K. S., editors, Image Processing: Machine Vision Ap-

plications VIII, volume 9405. SPIE.

Uijlings, J., Sande, K., Gevers, T., and Smeulders, A.

(2013). Selective search for object recognition. Intl.

Journal of Computer Vision, 104.

V. K. Bhatia, et. al (2010). Evaluation of Rationalization

of Minor Irrigation Statistics (RMIS) Scheme, Indian

Agriculture Statistice Research Institute, New Delhi.

Vakalopoulou, M., Karantzalos, K., Komodakis, N., and

Paragios, N. (2015). Building detection in very high

resolution multispectral data with deep learning fea-

tures. In 2015 IEEE Intl. Geoscience and Remote

Sensing Symposium (IGARSS).

Wagh, P., Das, D., and Damani., O. (2019). Well detec-

tion in satellite images using convolutional neural net-

works. In Proc., 5th Intl. Conf. on Geographical Infor-

mation Systems Theory, Applications & Management.

Yang, F., Chen, H., Li, J., Li, F., Wang, L., and Yan, X.

(2019). Single shot multibox detector with kalman

ﬁlter for online pedestrian detection in video. IEEE

Access, 7.

Yuan, J. (2018). Learning building extraction in aerial

scenes with convolutional networks. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence.

Yuan, J., Yang, H. L., Omitaomu, O. A., and Bhaduri, B. L.

(2016). Large-scale solar panel mapping from aerial

images using deep convolutional networks. In 2016

IEEE Intl. Conf. on Big Data.

Aiding Irrigation Census in Developing Countries by Detecting Minor Irrigation Structures from Satellite Imagery

215