An Oil Well Dataset Derived from Satellite-Based Remote Sensing

Abstract

Estimation of the number and geo-location of oil wells is important for policy holders considering their impact on energy resource planning. With the recent development in optical remote sensing, it is possible to identify oil wells from satellite images. Moreover, the recent advancement in deep learning frameworks for object detection in remote sensing makes it possible to automatically detect oil wells from remote sensing images. In this paper, we collected a dataset named Northeast Petroleum University–Oil Well Object Detection Version 1.0 (NEPU–OWOD V1.0) based on high-resolution remote sensing images from Google Earth Imagery. Our database includes 1192 oil wells in 432 images from Daqing City, which has the largest oilfield in China. In this study, we compared nine different state-of-the-art deep learning models based on algorithms for object detection from optical remote sensing images. Experimental results show that the state-of-the-art deep learning models achieve high precision on our collected dataset, which demonstrate the great potential for oil well detection in remote sensing.

1. Introduction

1.1. Background

Oil remains one of the largest energy sources globally despite the fact that renewable energy will be the fastest growing energy, according to BP’s 2019 Energy Outlook [1]. Petroleum production can affect soil and water (e.g., oil spills), which presents potential risks to the environment and public health. The monitoring of the number and geological distribution of oil wells, the status of oil exploitation, and an established energy early warning mechanism are essential, especially to the policy makers in energy resource management. Therefore, it is crucial to build an automatic oil well detection capability in order to address the challenges in oil well resources planning and environment monitoring. Traditional methods for oil well detection includes mainly relying on an on-site survey. Collecting geo-information by conducting surveys of oil wells is impractical and time consuming especially for large scale monitoring. Optical remote sensing techniques can periodically monitor oil wells by obtaining both spatial and temporal information, which show advantages of unrestricted national boundaries and objectivity of observation.

Oil is generally extracted from underground reservoirs based on drilling and pumping approaches. As underground infrastructure of oil wells are not seen, we use the aboveground infrastructure of oil well pump jacks (Figure 1) to detect the oil wells. Pump jacks are small targets in remote sensing, they are generally 6–12 m long, 2–5 m wide, and 5–12 m high, which may slightly vary depending on the model of the pump jack. In a 0.5 m resolution satellite image, a pump jack only has dozens of pixels with a small amount of target information. As can be seen in Figure 1, the land background of oil well pump jacks varies significantly including wetlands, forests, deserts, etc. Moreover, due to the distinctions in the pump jack installation, even the pump jacks of the same model show completely different shape in the images. Furthermore, the surrounding trees and other types of machinery will increase the difficulty in the detection of oil wells.

In recent years, the development of remote sensing technologies and the rapid improvement in deep learning-based object detection methods make it possible to monitor oil wells from optical remote sensing images. Remote sensing techniques have the advantages of short-term, non-contact, and wide area coverage, repetitive monitoring in earth observation, and are also able to monitor the objects objectively and periodically [2]. Object detection from optical remote sensing images determines objects of interests including vehicles, buildings, airplanes, ships, etc. It also plays a major role in a wide range of applications in environment monitoring and urban planning [3]. The aboveground infrastructure of oil wells shown in Figure 2 have sharp boundaries, which make oil wells independent from background environments; therefore, it is possible to detect oil wells from optical remote sensing images. In addition, the well site and the surrounding facilities of the oil wells, and the road connecting to the well site enable the oil well detection from multiple scales.

Satellite-based remote sensing images are taken from a high-altitude angle of view. There are several challenges in oil well detection. The oil wells are small targets in the remote sensing images (see Figure 3). They are often in the complex backgrounds and different in observation angles, which lead to the uncertainties in remote sensing monitoring and identification. Those challenges in oil well detection from remote sensing images are shown in Figure 3.

1.2. Related Work

1.2.1. Oil-Related Monitoring Using Remote Sensing Techniques

Most of the current relevant studies in the areas of oil-related monitoring are focused on oil spill detection, offshore drilling platforms, and oil tank detection. Oil spill monitoring has been an important topic in the past few decades as it is related to environmental monitoring or disaster monitoring [4,5,6]. For oil spill area detection, a variety of techniques have been used including visible light [7], infrared [8], near-infrared [9], ultraviolet [4], hyperspectral [10], image analysis [11], and satellite radar [12]. In recent years, there has been a significant increase in the research for monitoring of offshore oil resources. Liu et al. [13] used the contextual features extracted from the Defense Meteorological Program/Operational Line-Scan System (DMSP/OLS) night light data. Good recognition results had been achieved on extracting the spatial position information of the South China Sea oil and gas drilling platform by using the time series Landsat-8 Operational Land Imager (OLI) image and layered screening strategy. In recent years, most of the research interests are among the area of remote monitoring of the existing oil tank detection [14] from satellite images. In [15], a traditional machine learning algorithm with the Speeded up Robust Features (SURF) technique and Support Vector Machine (SVM) classifier has been used for oil tank detection. More recently, deep learning-based solutions have been applied in the detection of oil tanks to address the high computational complexity of remote sensing images [16]. In the detection of circular oil storage tanks, a circular support model was established based on the radiological symmetry characteristics of the object, which was verified by GeoEye-1 high-score images [17]. Zhang et al. [18] proposed a recognition algorithm based on deep environment features by using the convolutional neural network (CNN) model and SVM classifier to recognize oil tanks from complex backgrounds and achieved high recognition performance.

Some researchers [19] have used high-resolution satellite images and the Faster regional convolutional neural network (R-CNN) deep learning network (Zhang, Liu et al., 2018) to identify oil facilities and achieved good results. The facilities in well site represents the distribution of oil wells in a certain sense, but the well site can be water injection wells or multiple oil well deployments, which does not represent the true distribution of oil wells. To the best of our knowledge, currently there is no research about the oil well identification based on satellite remote sensing images.

1.2.2. Smart Oilfield

Digital oilfield [20] platform automates the workflow and monitoring the operations in order to reduce the cost and risks. Smart oilfield [21] is a new trend in the oilfield development based on the improvement in the previous digital oilfield. In the last few decades, with the ever-increasing big data, internet of things (IoT), and artificial intelligence (AI), the development of smart oil field has been further enhanced. Different sensors including pressure, density, and temperature sensor have been proposed and deployed to improve the safe and efficient oil pump extraction [22]. Some studies have focused on the exploration and identification of oil wells based on dynamometer card. For example, the electrical parameters from dynamometer card was used for recognition of the working condition of oil wells [23]. Sun et al., used CNN to construct the recognition model of oil well function diagram [24]. The recent remote sensing technologies can be used to remotely monitor the oil wells/oil equipment periodically, which can potentially provide additional data in the area of the smart oilfield.

1.2.3. Deep Learning in Remote Sensing

Recently, with the advancement in the area of optical remote sensing and deep learning algorithms, researchers have been working on detection of various ground targets including the detection of buildings [25,26], ships [27,28,29], airplanes [30,31]. Monitoring oil wells from remote sensing images becomes possible with the advancement of the object detection algorithms based on deep learning frameworks. The hand crafted features were used in traditional machine learning based object detection methods, which are difficult to be applied in massive data scenarios. More recently, a deep learning-based method has replaced the traditional machine learning method by utilizing (more) higher level and deeper features. Training of the deep learning models requires large scales of the dataset, and there are different datasets available for ships, flights and buildings [32,33,34]. In terms of applications related to oil, the detection of oil spills [35,36,37] and oil storage tanks [16] from remote sensing images based on deep learning frameworks received significant attention in recent years. In terms of oil well monitoring, very limited research has been done. Our previous work [38] detected the oil wells from remote sensing images using a Faster regional convolutional neural network (R-CNN) based model, but only a two-stage method has been used.

The use of deep learning based detection methods in remote sensing requires large scale dataset of labeled optical remote sensing images. For example, ImageNet [39], Common Objects in Context (COCO) [40], and pattern analysis, statistical modelling and computational learning (PASCAL) Visual Object Classes (VOC) [41] are widely used publicly available datasets, but none of these datasets provides images for oil wells. To the best of our knowledge, there are no available open oil well satellite image datasets. Therefore, in this work, we construct an oil well dataset from Google Earth Images, which aims to explore the feasibility of automatically detecting and dynamic monitoring of oil wells from remote sensing images. The dataset contains 1192 oil wells in 432 images, which vary in orientation and background. Moreover, we apply different state-of-the-art deep learning frameworks on our dataset to explore the best performance model. Our dataset will be publicly available online [42].

The remaining part of this paper is organized as follows. Section 2 introduces the construction of the dataset. Section 3 describes the state-of-the-art deep learning frameworks, which are used in this work. Section 4 describes the experimental results of applying the state-of-the-art models on our dataset and presents the accuracy of different models. Results are discussed in Section 5, and the conclusion is presented in Section 6.

2. Oil Well Dataset

2.1. Images Collection and Pre-Processing

We constructed a dataset named Northeast Petroleum University–Oil Well Object Detection V1.0 (NEPU-OWOD V1.0) of oil wells that contains the geographical coordinates and labels for each of the oil wells. Our selected research region for this work is in Daqing, China. Daqing is located in North China and is known as “Oil Capital of China”. Our dataset covers an area of 369 square kilometers as seen in Figure 4. The sample images in our dataset are 768 × 768 pixels, 768 × 1024 pixels and 1024 × 1024 pixels. Our dataset contains 432 images which contain 1192 oil wells from Google Earth Images with a very high resolution of 0.41 m per pixel. Figure 5 shows the distribution of the number of oil wells per image. It is noted that in most of the images, there are only 1 or 2 oil wells. In order to increase the diversity of the dataset, we included the images with the different backgrounds and various orientations of the oil wells (see Figure 6).

2.2. Images Annotation/Labeling

The images annotation has been done by local experts and researchers using an open-source annotation tool—RSLabel [43]. Each labeled oil well is saved to an XML file, which includes the geo-location coordinates of the oil wells, images sizes, and bonding boxes sizes, as shown in Figure 7.

3. Methods

State-of-the-art methods/algorithms: in order to evaluate our dataset, we chosen several state-of-the-art deep learning-based frameworks/models for comparison (see

Table 1. State-of-the-art deep learning models used in this work.

Model	Backbone
Faster R-CNN	ResNet-50-FPN
	ResNet-50-C4
	ResNet-50-DC5
	ResNet-101-FPN
	ResNet-101-C4
	ResNet-101-DC5
SSD	VGG16
YOLOv3	Darknet-53
RetinaNet	ResNet-50-FPN

). In general, the object detection algorithms are divided into two categories, including two-stage method and one-stage method.

3.1. Two-Stage Method

In the area of object detection based on deep learning neural networks, the convolutional neural network (CNN) lays the foundation. The first important approach in the two-stage method is regional convolutional neural network (R-CNN) [44], which initially extracts region proposal and computes features for each proposal based on CNN and then classifies each region. A number of improved approaches have been proposed based on R-CNN which include Fast R-CNN [45] and Faster R-CNN [46]. In this work, we mainly use Faster R-CNN based frameworks which firstly generate several regions of interests by using a Region Proposal Network and then send the proposed regions for object classification and bounding-box regression [47]. It is believed that the two-stage method can achieve higher accuracy rates but are typically slower compared with the one-stage method [48].

3.1.1. Faster R-CNN

In recent years, region proposal methods have been used in most of the object detection applications from the remote sensing images. Based on the first region-based convolutional neural networks (R-CNN) [44], new methods, including Fast R-CNN [45], and Faster R-CNN [46] have been proposed to improve the effectiveness and computation efficiency. As Faster R-CNN is better and faster, we only select Faster R-CNN for the second stage model and it has been used with different backbones (see Table 1). In this work, we used six different backbones for Faster R-CNN. Region Proposal Network (RPN) was proposed by Ren et al. [46] and it has been used in enormous application areas in object detection. For the objects from an input image, RPN is used to generate region proposals and an objectness score for each of the generated proposals. RPN consists of a classifier, which is used to determine the region proposal probability, and a regressor, which is used to regresses the proposal coordinates.

As shown in Figure 8, the deep convolutional networks extract the feature map from the input image, and Faster R-CNN uses convolutional network RPN to generate the region proposals. Then a Region of Interest (ROI) pooling layer is used to converting all the proposals to a fixed shape which is required by the final fully connected layer. ROI takes two inputs, from RPN and feature map, respectively, and flattens and passes them to two fully connected layers. Finally, two fully connected layers are applied to generate a prediction of the target and the best bounding box proposal.

3.1.2. Backbone Network

Backbone network works as a very important component in CNN based object detection frameworks [49]. In our evaluation, different existing backbone networks e.g., Residual Networks (ResNet) [50], Feature pyramid network (FPN) [51] with ResNet-50 and ResNet-101 have been utilized for feature extraction. In this work, for Faster R-CNN model we selected different combination of backbones including ResNet-50-FPN, ResNet-50-C4, ResNet-50-DC5, ResNet101-FPN, ResNet-101-C4, and ResNet-101-DC5. The evaluation of the performance of Faster R-CNN model with different backbone of oil well detections are presented in Section 4.

3.2. One-Stage Method

One-stage method (or single-stage method) is proposed in recent years to address the challenges in real-time object detection and it predicts straight from image pixels to bounding boxes and class probabilities utilizing only a single deep neural network. The most commonly used one-stage method includes You Only Look Once (YOLO) [52], Single Shot MultiBox Detector (SSD) [53], and RetinaNet [54]. These models usually run much faster compared with the two-stage method.

3.2.1. YOLO

YOLO is the first one-stage method based on convolutional neural network proposed by Redmon et al. [52]. It utilizes a unified architecture and predicts the bounding boxes location and target classification at the same time as a single regression problem by directly extracting the features from input images. Compared with the two-stage method, YOLO is faster. Different versions of YOLO have been proposed, e.g., YOLOv2 and YOLOv3, to improve the detection accuracy and shorten the running time. YOLOv2 [55] is the second version of YOLO and improves YOLO detection performance by incorporating a verity of techniques including batch normalization, high resolution classifier, convolutional with anchor boxes, dimension clusters, direct location predication, fine-grained features, and multi-scale training. YOLOv3 [56] is the third version of YOLO and has made further improvements on YOLO by utilizing a multi-label approach, which better models the data for complex dataset with overlapping labels. Moreover, three different scales were used in YOLOv3 for predicting bounding boxes based on the extracted feature maps and finally the last convolutional layers output three-dimensional (3D) tensor encoding bounding box, objectness, and class prediction. In this work, we used the YOLOv3 framework and the Darknet-53 as the backbone because Darknet-53 is better in detection accuracy and faster than ResNet, according to Redmon and Farhadi’s work [56]. The output predictions have been made at different scales. As in Figure 9, the YOLOv3 employs Darknet-53 as the backbone and then uses Darknet_conv2D_BN_Leaky (DBL) structures. As the basic element of YOLOv3, DBL structure is short for Darknet_conv2D_BN_Leaky which consists of a convolution layer, a batch normalization layer and a leaky ReLU layer [57]. Three different scale feature maps are used in the prediction stage.

3.2.2. SSD

SSD [53] is another one stage method that has been proposed in 2016 and it is a single-shot detector for multiple categories. It combines a feature extract and detection step based on a feed-forward convolutional network with a non-maximum suppression (NMS). Comparing with the two stage method, SSD removes the generation of the proposal region and incorporates the computation of features in just one stage. Compared with YOLO, SSD model adds several feature layers to the end of a base network [53]. As shown in Figure 10 below, SSD uses VGG16 to extract the feature map through using a Conv4_3 layer and a Pool5 layer. Eight more feature layers have been added after the base network (VGG16) in order to multi-scale feature maps for the detection. The use of multi-scale feature map significantly improves the detection accuracy.

3.2.3. RetinaNet

RetinaNet [54] is believed one of the best one-stage methods. It utilizes focal loss for the loss function, which addresses the class imbalance challenges among the one-stage method. The method chosen in this work uses ResNet-50-FPN backbone network as shown in Figure 11. In this framework, ResNet-50 is used to extract deep features. FPN is employed on top of the ResNet backbone to construct rich multi-scale feature pyramid from the input image. The probability of object presence at each spatial position and regresses the offset for bounding boxes are predicted by a classification and box regression subnet which is comprised of a fully convolutional network (FCN) attached to FPN.

4. Experimental Results

4.1. Training Details

In this work, we used our constructed dataset NEPU–OWOD V1.0, which contains 1192 oil wells from Google Earth Images with a very high resolution of 0.41 m per pixel. The size of images in our dataset varies from 768 × 768 pixels, 768 × 1024 pixels and 1024 × 1024 pixels. The training dataset included 345 images with 968 oil wells, which are randomly selected and the remaining is test dataset which included 87 images with 224 oil wells. All of the experiments were carried out on a server with Intel i9-9900KF CPU (3.60 GHz) and a NVIDIA GeForce RTX 2080Ti GPU (11264M).

Training Loss

The training loss changes are present in Figure 12 and Figure 13 for both two-stage methods and one-stage method, respectively. As shown in Figure 12, Faster R-CNN ResNet-101-DC5 (Figure 12d) is more robust with a faster convergence compared with other two-stage frameworks with different backbones used in this work. For one-stage model training loss (Figure 13), RetinaNet converges more quickly and has more robust training loss compared with YOLOv3 and SSD.

4.2. Evaluation Metrics

In this work, we mainly used several metrics, including precision, recall, F1 score, average precision (AP), and receiver operating characteristic (ROC) curve to evaluate different models proposed in Section 3.

4.2.1. Intersection over Union (IoU)

IoU, also known as Jaccard index, is one of the most commonly used evaluation metrics of object detection. IoU is computed as the Equation (1) below, it compares the similarities between the prediction and ground truth area.

$$IoU = \frac{\left|A\cap B\right|}{\left|A\cup B\right|}=\frac{\left|I\right|}{\left|U\right|},$$

(1)

where A and B are the prediction and ground truth bounding boxes, respectively. I is intersection Area and U is Union Area.

4.2.2. Precision, Recall, and F1 Score

Typically, we need to set a threshold of IoU to obtain the true positive (TP). When IoU > 0.5, the result is considered as TP.

$$Precision = \frac{TP}{TP+FP} ,$$

(2)

$$Recall = \frac{TP}{TP+FN} ,$$

(3)

$$\mathrm{F}1=\frac{2}{\frac{1}{\mathrm{Precision}}+\frac{1}{\mathrm{Recall}}}=2*\frac{\mathrm{Precision}*\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},$$

(4)

where TP is true positive, FP is false positive, and FN is false negative.

4.2.3. AP

AP combines recall and precision by computing the average precision over the recall and is defined as the area under the precision–recall curve as in Equation (5) below [58]. In object detection, the performance is usually evaluated by computing AP under different IoU thresholds [49]. AP50 and AP75, referring to the IoU threshold at 50% and 75%, are used, respectively. A higher AP value indicates a better performance.

$$AP = {{\displaystyle \int }}_0^{1}p\left(r\right)dr ,$$

(5)

4.2.4. McNemar’s Test

In order to evaluate if different state-of-art-models are significantly different from each other, McNemar’s test [59] is used. Equation (6) below shows the calculation of McNemar’s test statistic. The p-value threshold of 0.05 is used and the p-value lower than this threshold will indicate a significant difference between the two models.

$${\chi }^2 = \frac{{(b-c)}^2}{b+c} ,$$

(6)

where b is total number of test instances that the first model gets correct, but the second model gets incorrect, and c is total number of test instances that the first model gets incorrect, but the second model gets correct.

4.2.5. ROC Curve

ROC curve displays the sensitivity (TPR) on y-axis and (1-specitivity) (FPR) on x-axis graphically for varying cut-off points of the test values [60]. The area under the curve (AUC) is an important and effective measure for the performance of a model. A good model has an AUC close to 1, while a poor model has an AUC close to 0. Figure 14 shows the ROC curves for all nine state-of-the-art models used in this study. RetinaNet has the highest AUC (0.86) among all nine state-of-the-art models.

4.3. Experimental Results for State-of-Art Algorithms

4.3.1. Comparisons of Oil Well Detection Accuracy of Different State-of-the-Art Models

We compared nine different deep learning models for oil well detection. The comparison results are summarized in

Table 2. Oil well detection evaluation metrics (Precision, Recall, and F1 score) (Intersection over Union (IoU) = 0.5). The bold font indicates the best performance.

Model	Backbone	Precision	Recall	F1 Score	Training Time (Min)	Memory Cost (M)
Faster_R-CNN	R50-FPN	0.684	0.919	0.784	102	2131
	R50-C4	0.742	0.875	0.803	319	4379
	R50-DC5	0.738	0.897	0.810	171	5688
	R101-FPN	0.724	0.879	0.794	142	5200
	R101-C4	0.734	0.924	0.818	354	5211
	R101-DC5	0.764	0.928	0.838	196	7130
SSD	Darknet-53	0.807	0.674	0.734	33	1803
YOLOv3	VGG16	0.166	0.267	0.205	92	5107
RetinaNet	R50-FPN	0.645	0.892	0.749	147	6768

and

Table 3. Oil well detection average precision (AP) of nine state-of-art algorithms. The bold font indicates the best performance.

Model	Backbone	AP	AP50	AP75	Training Time (min)	Memory Cost (M)
Faster_R-CNN	R50-FPN	52.941	89.547	56.974	102	2131
	R50-C4	48.768	85.785	51.982	319	4379
	R50-DC5	50.230	89.469	48.969	171	5688
	R101-FPN	49.177	84.863	49.202	142	5200
	R101-C4	52.292	88.720	56.013	354	5211
	R101-DC5	50.034	87.408	52.935	196	7130
SSD	Darknet-53	43.298	85.853	36.103	33	1803
YOLOv3	VGG16	3.698	7.653	3.061	92	5107
RetinaNet	R50-FPN	55.529	99.795	55.317	147	6768

. The precision, recall and F1 score are presented for all of the deep learning models at IoU = 0.5 in Table 2. The results show that SSD model achieves the best precision with the shortest training time and lowest memory cost. Faster R-CNN with R101-DC5 backbone has the best performance in recall and F1-score. Faster R-CNN and RetinaNet models have relative high recall and F1-score. In Table 3, different AP have been calculated based different deep learning models. RetinaNet model has the best AP and AP₅₀ compared with all the other models and Faster R-CNN with R50-FPN backbone has the best AP₇₅. Figure 15 shows the samples of the detection results.

Table 4. p-value of McNemar’s Test for comparing different deep learning model performances. p-values under 0.05 indicate the two compared models are significantly different. (The bold font indicates models that were statistically different).

		Faster_R-CNN						SSD	YOLO v3	RetinaNet
McNemar’s Chi Squared (p-Value)		R50- FPN	R50-C4	R50- DC5	R101- FPN	R101- C4	R101- DC5	Darknet-53	VGG16	R50- FPN
Faster_R-CNN	R50-FPN
	R50-C4	0.025
	R50-DC5	1	0.008
	R101-FPN	0.180	0.285	0.083
	R101-C4	1	0.018	1	0.083
	R101-DC5	1	0.018	1	0.109	1
SSD	Darknet-53	<0.001	0.002	<0.001	<0.001	<0.001	<0.001
YOLOv3	VGG16	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001
RetinaNet	R50-FPN	0.593	0.074	0.637	0.371	0.637	0.637	<0.001	<0.001

presents the results of McNemar’s Test in order to explore whether the performance differences in different models are significant. Most of the Faster R-CNN models and RetinaNet model are statistically similar while Faster R-CNN ResNet50-DC5, SSD, and YOLOv3 show significantly differences with other models in model performance.

4.3.2. Comparisons of Oil Well Detection Accuracy in Different Background and Orientations

In order to further analyze the oil well detection performance under different background and at different orientations, we calculated the precision, recall, and F1-score for each background including bare land, trees, buildings, lakes, and also two different orientations of the oil wells which are horizontal and non-horizontal orientations.

Table 5. Total number of oil wells in the test dataset of different background and orientation.

	Bare Land	Trees	Buildings	Lakes	Horizontal	Non- Horizontal
Total number of the oil wells in the test dataset	148	36	30	10	157	67

shows the total number of the oil wells in different backgrounds and orientations in the test dataset.

Table 6. Oil well detection performance in different backgrounds and orientations (precision, recall, and F1 score (IoU = 0.5)) (The bold font indicates the best performance for each performance metric).

Model	Backbone	Performance Metric	Bare land	Trees	Buildings	Lakes	Horizontal	Non- Horizontal
Faster_R-CNN	R50-FPN	Precision	0.623	0.459	0.472	0.909	0.569	0.401
		Recall	0.926	0.944	0.833	1.000	0.924	0.821
		F1 score	0.745	0.618	0.602	0.952	0.704	0.539
	R50-C4	Precision	0.690	0.500	0.413	1.000	0.621	0.514
		Recall	0.919	0.861	0.633	1.000	0.885	0.851
		F1 score	0.788	0.633	0.500	1.000	0.730	0.640
	R50-DC5	Precision	0.688	0.443	0.477	0.909	0.621	0.504
		Recall	0.939	0.861	0.700	1.000	0.898	0.896
		F1 score	0.794	0.585	0.568	0.952	0.734	0.645
	R101-FPN	Precision	0.659	0.484	0.500	0.909	0.604	0.492
		Recall	0.912	0.833	0.733	1.000	0.885	0.866
		F1 score	0.765	0.612	0.595	0.952	0.718	0.627
	R101-C4	Precision	0.664	0.500	0.481	1.000	0.606	0.508
		Recall	0.946	0.889	0.833	1.000	0.930	0.910
		F1 score	0.780	0.640	0.610	1.000	0.734	0.652
	R101-DC5	Precision	0.688	0.443	0.477	0.909	0.646	0.541
		Recall	0.939	0.861	0.700	1.000	0.943	0.896
		F1 score	0.794	0.585	0.568	0.952	0.767	0.674
SSD	Darknet-53	Precision	0.735	0.452	0.536	1.000	0.703	0.530
		Recall	0.730	0.528	0.500	0.900	0.739	0.522
		F1 score	0.732	0.487	0.517	0.947	0.720	0.526
YOLOv3	VGG16	Precision	0.172	0.074	0.123	0.000	0.093	0.195
		Recall	0.311	0.194	0.233	0.000	0.178	0.478
		F1 score	0.221	0.107	0.161	0.000	0.123	0.277
RetinaNet	R50-FPN	Precision	0.604	0.394	0.424	0.909	0.569	0.401
		Recall	0.926	0.778	0.833	1.000	0.924	0.821
		F1 score	0.731	0.523	0.562	0.952	0.704	0.539
		Precision	0.614	0.417	0.434	0.838	0.559	0.454
Averages		Recall	0.839	0.750	0.666	0.878	0.812	0.785
		F1 score	0.706	0.532	0.520	0.856	0.659	0.569

shows the performance metrics for each of the background and orientations. It can be seen that the oil well detection accuracy varies with different backgrounds and orientations. The lake background has the best detection performance but there are only 10 images in the test dataset. Bare land background outperforms trees and buildings, which may due to the fact that the trees and buildings are more complicated. In terms of the orientation of the oil wells, it can be seen that the horizontal oil wells are easier to be identified than the non-horizontal oil wells.

5. Discussion

It is noted that many factors have influences on oil well detection, including the complex backgrounds and differences in observation angles, etc. Therefore, in our dataset, the oil wells in different backgrounds and orientations are included. Oil well clusters will be included in the future to further supplement this dataset.

We trained and tested our dataset on nine different state-of-the-art models, in which six models are two-stage detection models based on Faster R-CNN with various backbones and three models are one-stage detection models based on YOLOv3, SSD, and RetinaNet. The results presented in Section 4 show the SSD achieves the best performance with the highest precision (0.807) and SSD has the lowest training time and lowest memory cost in training. Faster R-CNN R101-DC5 model achieves the best recall (0.928) and F1 score (0.838). To further improve the model performance, we will focus on adjusting the structure of the current nine models to make them achieve better performance for oil well detection.

In the experimental results section, we also compared the model performance of the oil wells in different background and different orientations. For all of the models, the oil wells in the horizontal angle outperform the oil wells in the non-horizontal angle. It is known that the oil wells have the largest shadow when in the horizontal angle, which might work as a more distinguishable feature in training of the model. For oil well detection in all different backgrounds, the two stage model generally has better performance compared with the one stage method. For one stage models, RetinaNet outperforms SSD and YOLOv3 model. It is also noted that the background of lakes and bare land background have better performance than that of the trees and buildings. It may be due to the background of the trees and buildings make the image more complex.

6. Conclusions

In this work, we created an oil well remote sensing image dataset, NEPU–OWOD V1.0, which consists of 1192 oil wells in 432 images collected from Google Earth in Daqing, China. Several state-of-the-art models, including both two-stage methods and one-stage methods, were applied on this dataset. Nine state-of-the-art models were trained and tested on our dataset including Faster R-CNN based frameworks with five different backbones, SSD, YOLOv3, and RetinaNet. The experimental results show that deep learning-based methods are effective in the detection of the oil wells from the remote sensing images. The comparison between different methods and the results show that the deep learning based model is able to detect the oil wells effectively from optical remote sensing images. Moreover, this dataset will provide the opportunity for researchers to develop new object detection algorithms and improve the oil well detection in remote sensing. Researchers will be able to design new oil well detection algorithms utilizing our published dataset.