Well Detection in Satellite Images using Convolutional Neural Networks

Pratik Sanjay Wagh, Debanjan Das

∗

and Om P. Damani

Department of Computer Science & Engineering, IIT Bombay, Mumbai, India

Keywords:

Remote Sensing, Computer Vision, Object Detection, Convolutional Neural Networks, Machine Learning.

Abstract:

The Government of India conducts a well census every ﬁve years. It is time-consuming, costly, and usually

incomplete. By using transfer learning-based object detection algorithms, we have built a system for the auto-

matic detection of wells in satellite images. We analyze the performance of three object detection algorithms -

Convolutional Neural Network, HaarCascade, and Histogram of Oriented Gradients on the task of well detec-

tion and ﬁnd that the Convolutional Neural Network based YOLOv2 performs best and forms the core of our

system. Our current system has a precision value of 0.95 and a recall value of 0.91 on our dataset. The main

contribution of our work is to create a novel open-source system for well detection in satellite images and cre-

ate an associated dataset which will be put in the public domain. A related contribution is the development of

a general purpose satellite image annotation system to annotate and validate objects in satellite images. While

our focus is on well detection, the system is general purpose and can be used for detection of other objects as

well.

1 INTRODUCTION

Automation of well records promises to be a big

step in developing countries’ E-governance initia-

tives. Currently, the well census in India is conducted

every ﬁve years and in between, farmers are supposed

to contact the nearest revenue department ofﬁce to get

their land records updated with their well informa-

tion. This is rarely done, especially due to complex

ownership issues. Many government schemes offered

to farmers are contingent upon the assets held by the

farmer, e.g. to qualify for the drip irrigation sub-

sidy, and farmers are unable to take beneﬁt of these

schemes when records are not up-to-date.

Not just for individual farmers, but for the coun-

try as a whole, it is important to have good data on

groundwater use for planning and governance (Min-

istry of Water Resource, 2006a). For example, the

Central Groundwater Board estimates the country’s

net irrigated area and irrigation potential using the

well count and the counts of other ground and sur-

face water sources obtained through the minor irriga-

tion census. The process of estimating a watershed’s

groundwater development also requires the estimate

of groundwater use which uses the well count data.

There are several administrative and resource-

related challenges (Ministry of Water Resource,

∗

This author is also co-ﬁrst author.

2006b) in conducting the well census. This surveying

process is time-consuming, lengthy, and cost ineffec-

tive due to which the collected data is often incom-

plete and potentially corrupt. As a ﬁrst step towards

solving this problem, we exploit the recent advances

(Redmon et al., 2015; LeCun et al., 1999; Viola and

Jones, 2001; Dalal and Triggs, 2005; Albert et al.,

2017) in the area of applied machine learning and

computer vision, to automate the task of maintaining

well records. Later, we can overlay the cadastral maps

giving land ownership, with GIS maps with detected

wells, and classify well ownership by class, caste, re-

gions etc (Ministry of Water Resource, 2006c).

The task of automatic well detection is challeng-

ing since the actual real-life well looks quite different

from the wells present in satellite images. The con-

trast is depicted in Figure 2. This implies that object

detection methods that work well for real-life cam-

era images may not be suitable for object detection in

satellite images.

We observe that the wells in the Google Maps im-

ages are like dark circular shaped objects with a thin

white cyclic patch around it. A further challenge in

automatic well detection is that sizes of the wells in

developing countries are such that they generally look

very similar to tree or shadow of the trees (Figure 2),

and parts of the well maybe covered by the trees. This

poses challenge even for human annotators.

Wagh, P., Das, D. and Damani, O.

Well Detection in Satellite Images using Convolutional Neural Networks.

DOI: 10.5220/0007734901170125

In Proceedings of the 5th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2019), pages 117-125

ISBN: 978-989-758-371-1

117

Figure 1: Contrast between real-life well and wells in satel-

lite images (Maitra, 2011).

Figure 2: Similarity between well and shadow of a tree in a

satellite image.

While the analysis of satellite imagery is quite an

old ﬁeld, the application of deep learning in this area

is a new and emerging trend. Existing work in the area

focuses on detection of man-made objects in satellite

images, e.g, air-crafts(Wu et al., 2015), vehicles(Chen

et al., 2014) and oil tanks(Zhang et al., 2015). All of

these works follow a process pipeline approach. First,

the object is localized, then it is recognized, and then

the classiﬁcation happens. The binarized normed gra-

dients (Cheng et al., 2014) is used in (Wu et al., 2015)

to localize the air-crafts present in a satellite image

and then a Hybrid Deep Convolutional Neural Net-

work is used for feature extraction, and for classiﬁca-

tion support-vector machines (Vapnik, 1999) is used.

Similarly, in (Zhang et al., 2015), the object of in-

terest is oil-tanks, and hence the elliptical or circular

shapes in an image are detected using ellipse and line

segment detector. After object localization, they use

Histogram of Oriented Gradients (Dalal and Triggs,

2005) and Convolutional Neural Networks methods

for feature extraction.

All of these methods require large training sets.

In the area of urban planning where training data is

scarce, the concept of transfer learning is used in (Al-

bert et al., 2017) for object classiﬁcation and feature

extraction.

In the context of general object detection algo-

rithms, we ﬁrst worked with the basic Convolu-

tional Neural Network architectures (LeCun et al.,

1999), HaarCascade (Viola and Jones, 2001) and

Histogram of Oriented Gradients (Dalal and Triggs,

2005) with support-vector machine (Vapnik, 1999)

methods. However, we achieved limited success to-

wards well detection with these approaches.

Then, inspired by the success of transfer learn-

ing approach in urban planning context (Albert et al.,

2017), we decided to experiment with the transfer

learning technique in the context of rural planning.

We experimented with the popular single shot object

detection algorithms You Only Look Once (YOLOv2

(Redmon and Farhadi, 2016)). In this work, the prob-

lem of object detection is framed as a regression prob-

lem where it outputs the probable bounding box co-

ordinates and the class probabilities for each bound-

ing box. YOLO uses Deep Convolutional Neural Net-

works for image recognition and bounding box draw-

ing. In contrast to the pipeline approaches, it performs

the object detection in the single shot, which makes it

a lot faster than other object detection algorithms.

In the literature, several authors lament the

scarcity of good quality data for object detection in

satellite images and manually collect data from dif-

ferent sources. For our well detection problem also,

there is no existing satellite image dataset with wells

identiﬁed in it. As we collaborated with members of

the Rural Development Department for data collec-

tion, we realized the need for a proper system to an-

notate and validate wells in the satellite images. To

our surprise, no such system was available in the pub-

lic domain. Hence, we decided to develop a general

purpose satellite image annotation system (SIAS) us-

ing which one can annotate and validate objects of

any type in satellite images.

The contribution of our work is two-fold. The

main contribution of our work is to create a novel

open-source transfer learning based system for well

detection in satellite images and create an associated

dataset to be put in the public domain. A related

contribution is the development of a general purpose

satellite image annotation system to annotate and val-

idate objects in satellite images. While our focus is

on well detection, the system is general purpose and

can be used for detection of other objects as well.

GISTAM 2019 - 5th International Conference on Geographical Information Systems Theory, Applications and Management

118

In Section 2, we discuss our SIAS system. Section

3 describes the object detection algorithms which we

have used, in Section 4 we discuss about the dataset,

performance metric we used and the training proce-

dure we followed and the last section contains the

conclusions.

2 SYSTEM ARCHITECTURE

In addition to experimenting with machine learning

algorithms, it is important to build a system that is in-

tuitive to use for end users. In our case, end users will

be government ofﬁcials who may not be very com-

puter savvy. Hence we developed a web-based ap-

plication that can be used across the internet. Further,

the same system should be capable of taking feedback

from the user where the user can validate or invalidate

the detected well. A user should also be able to mark

an undetected well. In fact, this feedback should ﬂow

into the system and get added to the training set. This

way the training set will keep growing over time as the

system gets used. We call our system SIAS - Satellite

Image Annotation System.

The architecture diagram of the whole system is

given in Figure 3.

2.1 User Interface

We follow the Software As A Service (SaaS) model.

At the backend, we use Flask (Ronacher, 2010) which

is a micro-web framework. The frontend is developed

with HTML, CSS, and JavaScript. The user interface

is shown in Figure 4.

2.1.1 Frontend

The adoption of the system critically depends on the

user-friendliness of the frontend. Various features of

frontend are listed below.

• It contains a search bar and a map window where

one can search for a place and the resulting satel-

lite image is shown in that window.

• There is an optional ﬁeld, using which we can se-

lect the model we want to use. The available mod-

els are YOLOv2, tinyYOLOv3, Basic Convolu-

tional Neural Network model, the Histogram of

Oriented Gradients-support-vector machines and

HaarCascade.

• After searching for a place and selecting the

model, clicking on ”Predict” button shows the

predicted wells in that area with the help of

bounding boxes. These bounding boxes are su-

perimposed on the satellite image.

• Note that user can select an arbitrarily large area

to predict. We subdivide the selected area into

multiple blocks of size 640 × 640. Each of these

blocks is then given as input to the selected model.

Currently, if a well falls on the boundary of two

sub-blocks then it is not detected. In the future,

we plan to solve this problem by creating differ-

ent partition sets of the same area and merging the

results from these partition sets.

• An important feature of the system is that it pro-

vides the user ability to give feedback. Each of the

predicted boxes can be marked as valid or invalid

by the user. This way the training set will keep

growing over time as the system gets used.

• The user can also label any object of interest in the

image by selecting and marking the correspond-

ing area.

2.1.2 Backend

The backend acts as a bridge between the object de-

tection models and frontend. Images from frontend

come in the form of a base64 string. In the backend,

it is decoded and converted into PNG images and the

model is applied to it. The output image is then sent

back to the frontend.

3 OBJECT DETECTION

ALGORITHMS

We experimented with various object detection meth-

ods and found that transfer learning based YOLO

model performs the best. We next describe the meth-

ods tried.

3.1 Basic Convolutional Neural

Network

In this model we have used CNN as a feature extrac-

tor and multi-layer perceptron (MLP) as a classiﬁer.

We have created a two-layered Convolutional Neural

Network (CNN) model. The description of the model

is given as:

• The ﬁrst convolutional stage is with rectiﬁed lin-

ear unit (ReLU) activation function followed by

a max-pooling layer. This layer has 20 convolu-

tional ﬁlters, each one of them has a size of 5 × 5.

The output dimension is the same as that of the

input shape i.e, 64 × 64. As this is the ﬁrst layer

in our model, we have to deﬁne its input shape

i.e, (1,64,64). The max-pooling operation applies

a sliding window that slides over the image and

Well Detection in Satellite Images using Convolutional Neural Networks

119

Figure 3: SIAS system architecture.

Figure 4: User interface of the SIAS system.

takes the maximum value of the pixel present in

that region.

• The second convolutional layer is also followed

by rectiﬁed linear unit activation function and a

max-pooling layer. Now, we increase the number

of convolutional ﬁlters to 50 from 20 in the previ-

ous layer.

• Then we have a standard ﬂattening layer and a

dense layer of 500 neurons. This ﬂattening layer

ﬂattens the last 3D feature block of the CNN and

provides it to the dense layer. This dense layer

acts as a hidden layer. The output layer of the

MLP has softmax function as activation function,

which can be deﬁned as:

P(y = j|X) =

·w

∑

k=1

·w

GISTAM 2019 - 5th International Conference on Geographical Information Systems Theory, Applications and Management

120

where K is the number of classes and P(y = j|X)

is the probability of the input being in class j

given it’s feature vector X.

3.2 HaarCascade

Introduced in 2006, it has been very successful at

face detection problem(Viola and Jones, 2001). Since

then it has been used for various other object detec-

tion problems which use non-satellite images. It uses

the Haar-like feature for detecting an object. It has

introduced an integral image for efﬁcient and quick

calculation of the sum of pixel values in a rectangular

subset of the given image, thus lowers the computa-

tion time. It uses AdaBoost classiﬁer for classifying

the positive and negative feature.

3.3 HOG

Histogram of Oriented Gradient (HOG)(Dalal and

Triggs, 2005) is a feature descriptor method. This

method uses the orientation of an image as a feature

for detecting an object thus generalizes well for the

objects which have a ﬁxed shape. It ﬁrst divides the

image into small connected blocks which consist of

cells and then computes the histogram of the gradient

for each pixel within the cell. Then it discretizes each

cell into angular bins according to the gradient ori-

entation. Each cell is combined to form blocks, these

blocks are normalized and used as a feature descriptor

which is used with support-vector machine (Vapnik,

1999) for object detection.

3.4 YOLOv2

You Only Look Once (YOLOv2) (Redmon and

Farhadi, 2016) introduced in late 2016 gave a mean

Average Precision (mAP) of 76.8 on VOC 2007

dataset, is an improved version of the YOLOv1 (Red-

mon et al., 2015). Unlike the traditional network, it

uses a single neural network for object classiﬁcation

and detection. It divides the whole image into grids

of small boxes and each box is responsible for de-

tecting and predicting the class. The whole process

occurs in one pass thus a global context of an im-

age is available for detecting the object in an image.

It uses batch normalization, high-resolution classiﬁer,

convolution with anchor boxes, direct location predic-

tion, and multiscale prediction.

It uses Darknet-19 as base model which consists

of 19 convolutional layers and 5 max-pooling layers

which requires 5.58 billion operations to process an

image thus making it faster.

3.5 YOLOv3

YOLOv3 (Redmon and Farhadi, 2018) is the succes-

sor approach for the YOLOv2. It was introduced in

2018 and incorporated some of the drawbacks present

in YOLOv2 such as upsampling, skip connections and

residual blocks. It works in a similar way as the

YOLOv2 with some variation in the architecture. It

uses Darknet-53 as the base model which consists of

106 convolutional layers making it slow in detection.

The accuracy is increased due to the use of 1x1 de-

tection kernel at three different scale. It uses 9 an-

chor boxes instead of 5, 3 at each different scale. This

anchor box is arranged in the descending order of di-

mensions and thus detects the 10x more no. boxes

than YOLOv2.

4 EXPERIMENTAL RESULTS

As mentioned before, different object detection meth-

ods require different dataset for training and follow

different training and testing procedure. We next de-

scribe below the dataset which we used for each ob-

ject detection method, performance measure, and the

training procedure which we followed.

4.1 Dataset Description

The uniqueness of the problem resulted in data

scarcity. Well detection is a completely new prob-

lem that does not have any dataset as per the best of

our knowledge. The dataset we used for each object

detection algorithm is described below :-

• Basic Convolutional Neural Network: We have

prepared a different dataset of ground truth im-

ages of well and non-well images. We have man-

ually collected the coordinates of the wells and

downloaded the image of that coordinate from

Google Earth at zoom level of 19. We have col-

lected around 769 well images and 700 non-well

images of size 64 × 64.

For training the model, 500 well images and 500

non-well images are used. And for testing the

model 269 well images and 200 non-well images

have been used.

• HaarCascade: We trained Haarcascade for a dif-

ferent dataset as it requires two sets of data i.e

positive data which only contains the well image

and the negative data which do not contain the

well image. Thus we have manually collected the

coordinates of the wells and non-wells (which is

anything except well) and downloaded the image

Well Detection in Satellite Images using Convolutional Neural Networks

121

of that coordinate from Google Earth at a zoom

level of 19. We trained using 1721 positive im-

ages which has a dimension of 64 x 64 and 3035

negative images which has a dimension of 114 x

114.

• YOLOv2 and tinyYOLOv3: We developed a

dataset of 1068 images of resolution 640 × 640

from Google Maps Static API at a zoom level of

19. Each image contains multiple wells. We had

drawn ground truth bounding boxes around the

wells in an image using BBox-Label-Tool (Qiu,

2017).

We collected the images from 20 different districts

of Maharashtra state. Out of which training data

consists of 535 images from 10 districts and test-

ing data consists of 403 images from other 10 dis-

tricts. We randomly extracted 130 images from

the districts of Maharashtra state as a validation

dataset. We trained the model for 535 images and

validated it on a set of 130 images. After vali-

dation we found that out of 130 images, 100 im-

ages were such in which most of the wells present

were not getting detected, so we added those 100

images to the training dataset, thus totalling the

training dataset to 635 images and then retrained

our model for this set.

• HOG: Thus to compare HOG with YOLO mod-

els we extracted the wells present in the images of

training dataset, which we described in the YOLO

model section consisting of 635 images, at a di-

mension of 150 x 150 which totals to 1022 posi-

tive images. Then we annotated the images using

imglab tool and trained HOG on the newly created

dataset along with the annotated ﬁle.

4.2 Performance Measure

To evaluate the performance of our model we used

some metrics which helped us to ﬁnd the best object

detection algorithm for our model. Description re-

lated to our context along with the formula is listed

below.

• Precision: It is the fraction of the wells detected

among all the detections in an image. The formula

for calculating the precision is given below.

Precision =

True Positives

True Positives + False Positives

• Recall: It is the fraction of the wells detected

among all the wells present in an image. The for-

mula for calculating the recall is given below.

Recall =

True Positives

True Positives + False Negatives

• F1 Score: It is the harmonic mean of Precision

and Recall and the formula for the F1 score is

given below.

F1score = 2 ×

Precision × Recall

Precision + Recall

4.3 Training

Different methods follows different training proce-

dure thus we mentioned one by one training proce-

dure which we followed for each method :-

• Basic Convolutional Neural Network The results

of deploying the above mentioned trained model

on 469 test images are given below:

– Test accuracy Score: 0.92

– Correctly classiﬁed images: 435

– True Negatives: 211

– True Positives: 224

– False Positives: 19

– False Negatives: 15

– Precision: 0.92

– Recall: 0.93

Although the results of the model looks promis-

ing on this dataset but it performed poorly on

the other dataset, i.e, dataset from a different ge-

ographical region. This model takes images of

64 × 64 resolution and predicts whether it is well

or not. Thus deploying this model on 640 × 640

images required proper object localization proce-

dures. Upon applying sliding window approach,

we found out that the number of false positive

were very high i,e, the precision was very low but

the true positive rate or the recall was high. Thus

we dropped this method from comparison. Some

results are shown in ﬁgure 5.

Figure 5: Results of Basic CNN on 640 × 640 images.

• HaarCascade We trained using 1721 positive im-

ages and 3035 negative images and the results ob-

tained are shown in Figure 6. As can be seen

in Figure 6 the results obtained were terrible and

GISTAM 2019 - 5th International Conference on Geographical Information Systems Theory, Applications and Management

122

improving them will require a much larger num-

ber of positive and negative images which was not

feasible for us. Thus we dropped this algorithm

from comparison.

(a) (b)

Figure 6: (a) and (b) are results obtained from Haarcascade

method.

• HOG

We trained HOG using dlib on a machine which

has 8 GB RAM. The training was conﬁgured with

a learning rate of 0.01, epsilon value of 5 and slid-

ing window of 83 x 77 pixel.

The results obtained can be seen in Figure 7

which are promising thus we used this method for

comparison with other models.

(a) (b)

Figure 7: (a) and (b) are results obtained from HOG

method.

• YOLOv2

For training YOLOv2 and using it for detec-

tion we used darknet framework. The training

was performed on the standard conﬁguration of

YOLOv2-VOC (yolov2-voc.cfg) it has a learning

rate of 0.001, the batch size of 64 and subdivisions

of 8. We used a weights ﬁle already pretrained

on Imagnet dataset (Russakovsky et al., 2015) and

trained on GPU (Nvidia Titan XP) for 20000 iter-

ations. The object detection was performed on the

default threshold value of 0.25 or more.

The results obtained are shown in Figure 8. This

result obtained is the best result achieved as com-

pared to methods which we discussed till now.

(a) Result-1 (b) Result-2

Figure 8: Results from YOLOv2 obtained from the weight

ﬁle trained for 20000 iterations.

• tinyYOLOv3

We used the Darknet framework for training and

detection. The training was performed on the

standard conﬁguration of yolov3-tiny which is

a constrained network of YOLOv3 (yolov3-tiny-

obj.cfg) it has a learning rate of 0.001, the batch

size of 64 and subdivisions of 8 We used a weights

ﬁle already pretrained on Imagnet dataset (Rus-

sakovsky et al., 2015) and trained on GPU (Nvidia

Titan XP) for 20000 iterations. The object detec-

tion was performed on the default threshold value

of 0.25 or more. Similar results as YOLOv2 were

obtained for the images in Figure 8.

4.4 Retraining Surprise

One would think that inclusion of an image set in

training set will ensure that the trained model will give

high detection accuracy on that image set. However,

while this holds true for YOLO, surprising this is not

true for other object detection algorithms that we ex-

perimented with.

We validated the weights ﬁle obtained from train-

ing for 130 well images extracted randomly from any

district present in Maharashtra state for YOLO mod-

els. After validation, we found that in 100 images

from the validation dataset some of the wells present

were not getting detected thus we added those 100 im-

ages to the training dataset and retrained the network

for 20000 iterations. The results obtained before and

Well Detection in Satellite Images using Convolutional Neural Networks

123

after retraining can be seen in Figure 9.

(a) (b)

Figure 9: Figure (a), (c) are results obtained from YOLOv2

before validation and Figure (b), (d) are corresponding re-

sults after validation and retraining on YOLOv2.

The results show that if we retrain YOLO model

by including images in which wells were not getting

detected then the retrained model is able to detect the

similar wells which were not getting detected earlier.

This might be because the wells which were not get-

ting detected vary totally from the images for which

we have trained our model.

For the other object detection methods not based

on transfer learning, even after retraining the model

by including such images in training dataset, simi-

lar wells present in other images were not getting de-

tected. We have not been able to ascertain the reason

for the same.

4.5 Results

For testing we extracted 403 well images from 9 dis-

tricts of Maharashtra which includes Solapur, Kolha-

pur, Hingoli, Ratnagiri, Beed, Aurangabad, Chandra-

pur, Jalna, and Satara.

Then we tested the weights ﬁle of YOLOv2 and

tiny YOLOv3 obtained from 20000 iterations and

HOG using the svm ﬁle obtained after training on the

testing dataset.

The results can be seen in Table 1.

Thus the obtained precision and recall value on the

testing dataset for YOLOv2 is 0.95 and 0.91, for tiny

Table 1: Performance comparison of YOLOv2, tiny

YOLOv3 and HOG model.

Metric Model

Y OLOv2 tiny Y OLOv3 HOG

True Positive 592 529 226

False Positive 29 9 12

False Negative 58 100 430

Table 2: Precision-Recall trade-off between YOLOv2, tiny

YOLOv3 and HOG.

Model Precision Recall F1 Score

Y OLOv2 0.95 0.91 0.92

tiny Y OLOv3 0.98 0.84 0.90

HOG 0.95 0.34 0.51

YOLOv3 is 0.98 and 0.84 and for HOG is 0.95 and

0.34 respectively.

The precision-recall trade-off between these algo-

rithms is shown in Table 2. Based on these results, our

system uses YOLOv2 by default, but we provided the

option to user to chose tiny YOLOv3, HOG or other

algorithms if they wish.

5 CONCLUSIONS

We have built an open source system for well detec-

tion in satellite images and an associated annotation

system to annotate and validate identiﬁed objects in

satellite images. The advantage of the annotation sys-

tem is that we can use it to continuously grow our

dataset. We plan to use data augmentation technique

to create a larger dataset for transfer learning. Also,

YOLO was pre-trained on ImageNet (Russakovsky

et al., 2015) dataset which is very different from satel-

lite images. For the future work, we want to train

YOLO on some satellite image dataset, e.g, xView

dataset (Experimental, 2018). On top of that, we can

train it on our enhanced well dataset. Also, further

experimentation with several single shot object de-

tection algorithms has to be done along with transfer

learning to develop a robust system. Our aim would

be to generalize this satellite image object detection

procedure as much as possible to several classes of

objects.

REFERENCES

Albert, A., Kaur, J., and Gonzalez, M. C. (2017). Using

convolutional networks and satellite imagery to iden-

tify patterns in urban environments at a large scale.

CoRR, abs/1704.02965.

GISTAM 2019 - 5th International Conference on Geographical Information Systems Theory, Applications and Management

124

Chen, X., Xiang, S., Liu, C., and Pan, C. (2014). Vehicle

detection in satellite images by hybrid deep convolu-

tional neural networks. IEEE Geoscience and Remote

Sensing Letters, 11(10):1797–1801.

Cheng, M., Zhang, Z., Lin, W., and Torr, P. (2014). Bing:

Binarized normed gradients for objectness estimation

at 300fps. In 2014 IEEE Conference on Computer

Vision and Pattern Recognition, pages 3286–3293.

Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-

dients for human detection. In 2005 IEEE Computer

Society Conference on Computer Vision and Pattern

Recognition (CVPR’05), volume 1, pages 886–893

vol. 1.

Experimental, D. I. U. (2018). Diux xview 2018 detection

challenge (http://xviewdataset.org).

LeCun, Y., Haffner, P., Bottou, L., and Bengio, Y. (1999).

Object recognition with gradient-based learning. In

Shape, Contour and Grouping in Computer Vision,

pages 319–, London, UK, UK. Springer-Verlag.

Maitra, M. K. (2011). Frequently asked questions

(faq) on groundwater - understanding the basics

(http://www.indiawaterportal.org/articles/frequently-

asked-questions-faq-groundwater-understanding-

basics).

Ministry of Water Resource, G. o. I. (2006a).

Census of minor irrigation schemes

(http://micensus.gov.in/censusmisch.html).

Ministry of Water Resource, G. o. I. (2006b).

Methodology of minor irrigation census

(http://micensus.gov.in/methodology.html).

Ministry of Water Resource, G. o. I. (2006c). National level

report - dugwell (http://micensus.gov.in/dugnat.html).

Qiu, S. (2017). Bbox-label-tool

(https://github.com/puzzledqs/bbox-label-tool).

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

(2015). You only look once: Uniﬁed, real-time object

detection. CoRR, abs/1506.02640.

Redmon, J. and Farhadi, A. (2016). YOLO9000: better,

faster, stronger. CoRR, abs/1612.08242.

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

improvement. CoRR, abs/1804.02767.

Ronacher, A. (2010). Flask (http://ﬂask.pocoo.org/).

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). Ima-

genet large scale visual recognition challenge. Int. J.

Comput. Vision, 115(3):211–252.

Vapnik, V. N. (1999). An overview of statistical learn-

ing theory. IEEE Transactions on Neural Networks,

10(5):988–999.

Viola, P. and Jones, M. (2001). Rapid object detection us-

ing a boosted cascade of simple features. In Proceed-

ings of the 2001 IEEE Computer Society Conference

on Computer Vision and Pattern Recognition. CVPR

2001, volume 1, pages I–I.

Wu, H., Zhang, H., Zhang, J., and Xu, F. (2015). Fast air-

craft detection in satellite images based on convolu-

tional neural networks. In 2015 IEEE International

Conference on Image Processing (ICIP), pages 4210–

4214.

Zhang, L., Shi, Z., and Wu, J. (2015). A hierarchical oil

tank detector with deep surrounding features for high-

resolution optical satellite imagery. IEEE Journal of

Selected Topics in Applied Earth Observations and

Remote Sensing, 8(10):4895–4909.

Well Detection in Satellite Images using Convolutional Neural Networks

125