Storage tank detection in remote sensing images based on circular bounding boxes and large selective kernel Authors: Yu Liu, Yong Wan, Weimin Huang, Zihao Zhan Published: 2025 ABSTRACT AbstractAccurate storage tank detection in remote sensing images is vital for monitoring methane emissions, a potent greenhouse gas, from the oil and gas industry. Existing methods, such as traditional geometric and spectral feature-based approaches, suffer from high false detection rates due to background variations and imaging conditions, while deep learning models like YOLO series and EfficientDet struggle with small objects, multi-scale features, background interference, and regression sensitivity, leading to missed detections and false positives. This study introduces a novel method integrating circular bounding boxes and a Large Selective Kernel (LSK) to address these gaps. Circular bounding boxes, aligned with storage tanks’ typical circular shape, stabilize Intersection over Union (IoU) for small objects, while LSK dynamically adjusts the receptive field to leverage contextual information effectively. Implemented on a YOLO-v10 framework and evaluated on a comprehensive dataset comprising DIOR, NWPUU_RESISC45, NWPU VHR-10, TGRS-HRRSD, and a self-built dataset (totaling 3568 images and 46075 storage tanks), our approach achieved a precision of 0.911, recall of 0.902, and mean Average Precision (mAP@0.5) of 0.931. These results represent improvements of 2.0% in precision, 2.7% in recall, and 1.8% in mAP@0.5 over the state-of-the-art YOLO-v10 baseline, offering a robust tool for pinpointing methane emission sources and supporting environmental sustainability efforts in the oil and gas sector.Keywords: Circular bounding box, Methane, Storage tanks, Remote sensing, Target detection.Subject terms: Energy infrastructure, Information technology Introduction Methane, a potent greenhouse gas with 25 times the global warming potential of CO₂, significantly contributes to climate change. Human activities account for 60% of atmospheric methane emissions, of which the oil and gas industry is a major contributor1,2. Specifically, oil and gas storage tanks are among the primary sources within this sector, responsible for up to 94.1% of methane emissions in the oil and gas industry3. Therefore, accurately identifying storage tanks is critical for pinpointing emission sources and supporting methane mitigation efforts. However, wide and dispersed geographic distribution of storage tanks, combined with the fact that the coordinates of many tanks are not recorded, means that simple statistical screening is insufficient for accurately locating all storage tanks. With the increasing resolution of remote sensing technology in recent years4, along with its inherent advantages of wide coverage and fast data acquisition, remote sensing has become an essential tool for detecting and identifying oil and gas storage tanks5. In addressing the challenge of detecting storage tank targets, several studies have employed a combination of conventional image processing methodologies, including morphological processing, Hough transform, clustering, and minimum distance classifiers. These methodologies have been utilized to leverage spatial and spectral characteristics for the purpose of high-resolution remote sensing image detection of storage tank targets. However, storage tanks are typically located in complex backgrounds. Traditional geometric and spectral feature-based methods have limited performance under background and imaging condition variations, resulting in a high false detection rate in storage tank detection6, primarily due to their inability to handle variations in lighting, shadows, and similar circular structures. Therefore, deep learning-based methods have progressively emerged as the prevailing approach, exemplified by the YOLO series (e.g., YOLOv5, YOLOv7) and EfficientDet, which have garnered extensive application. These methods incorporate attention mechanisms (e.g., CBAM, SimAM, multi-head attention), feature enhancement modules (e.g., C3, CARAFE), and small object detection heads to enhance detection capabilities in complex backgrounds of remote sensing images7,8. Deep learning methods perform better in detecting storage tanks in complex backgrounds, but further optimization is needed to improve detection accuracy and efficiency. One of the key challenges is the presence of many small objects in remote sensing images, and small object detection has long been one of the most significant challenges in storage tank detection methods9,10. Small object detection continues to face many difficulties due to the limited size of small objects, fewer visual and geometric cues, and the lack of large-scale datasets for small objects. Some of the challenges include the complexity of multi-scale feature learning11, the impact of background interference and dense targets12, the absence of contextual information13, and the sensitivity of the regression branch14. For storage tanks, remote sensing images typically exhibit multi-scale target characteristics, where both large and small storage tanks may be present simultaneously. Suppose the focus is solely on improving the detection performance for small objects (e.g., by lowering the Intersection over Union (IoU) threshold). Then in such a case, it may decrease the detection model’s overall performance15. Therefore, an efficient multi-scale object detection method is required to accurately identify large and small storage tanks in remote sensing images. Storage tanks generally appear as circular objects from the overhead perspective of remote sensing imagery. However, many human activity traces on the ground appear circular when viewed from above, which can lead to false identification of these features as storage tanks. This issue arises because conventional object detection methods mainly focus on the pixels within the target boundary while neglecting the surrounding background pixel information, which could be useful. Contextual information plays a crucial role in improving detection accuracy by providing additional cues to distinguish targets from similar backgrounds, reducing false positives. The receptive field must be expanded to effectively incorporate background information, typically increasing the convolutional kernel size. However, this approach can result in losing small object features16,17. Therefore, it is necessary to dynamically adjust the receptive field based on the scene requirements, reducing the false detection rate of the object detection method. Recent advances in remote sensing object detection include oriented bounding box methods18 and contextual enhancement techniques19, yet challenges like regression sensitivity and background noise persist. To solve the problems of false detections and missed detections in storage tank target detection tasks in remote sensing images, this paper proposes a detection method based on circular bounding boxes combined with a large selective kernel (LSK). Circular bounding boxes better align with the morphological characteristics of storage tanks in remote sensing images, and theoretical analysis shows that circular bounding boxes provide greater IoU stability during slight positional shifts, which helps improve the recall of small objects. A circular bounding box detection framework is introduced to solve traditional rectangular bounding boxes’ regression branch sensitivity issue. Meanwhile, the LSK adaptively selects convolution kernel sizes, dynamically adjusting the model’s receptive field to efficiently utilize background information in remote sensing images, thereby improving detection accuracy. By integrating these two methods, we aim to significantly enhance oil and gas storage tank detection performance, providing robust technical support for identifying potential greenhouse gas emission sources. The major contributions of this work are: (1) addressing IoU instability for small circular objects, thereby reducing missed detections and improving recall in multi-scale scenarios; (2) enabling adaptive receptive field adjustment to mitigate background interference, enhancing precision in complex environments without sacrificing efficiency. These advancements target the unique gaps in storage tank detection, such as regression sensitivity and improper contextual utilization, distinguishing our work from general object detection methods. However, limitations include the method’s specificity to circular objects, potentially limiting generalization to non-circular targets, and increased computational complexity from LSK, which may demand more resources. Future research could explore shape-adaptive bounding boxes and optimize LSK efficiency. This paper is organized as follows. Section 2 presents an overview of the related works, focusing on small object detection in remote sensing images and the dynamic adjustment of receptive fields. Section 3 outlines the problem statement, identifying the key challenges in storage tank detection. Section 4 describes the proposed methodology, including the use of circular bounding boxes and the LSK. Section 5 provides details about the experimental setup, datasets used, and results obtained. Section 6 provides a comprehensive discussion of the experimental results. Finally, Sect. 7 concludes the paper with a summary of key results and future directions for research. Related works This section provides a concise overview of prior research relevant to storage tank detection in remote sensing images, focusing on three key areas: traditional and deep learning methods for storage tank identification, advancements in small object detection, and techniques for dynamically adjusting receptive fields. These areas highlight the evolution of detection strategies, their limitations in addressing small, circular targets like storage tanks, and the motivation for our proposed enhancements, which build upon these foundations to improve accuracy and robustness. Storage tank target detection from remote sensing images Storage tank detection from remote sensing images has made significant progress with the advancement of deep learning. Traditional methods are simple and computationally efficient, but they rely on geometric and spectral features. When local target distribution patterns or imaging conditions change, these methods become limited, leading to high false positives in different backgrounds. Hierarchical residual learning branches have shown effectiveness in identifying storage tanks in large-scale, high-resolution images with strong generalization ability20. However, it is computationally intensive. Improved models like EfficientDet7 and Faster R-CNN21 have significantly enhanced detection accuracy, but they still fail to resolve the issue of slow inference speeds caused by computational intensity. By extracting contour shape cues and using sparse reconstruction to generate saliency maps, storage tanks can be accurately located while minimizing interference from high-contrast backgrounds22. However, given the effects of satellite angle and occlusion, this method is not capable of precisely identifying all tanks within complex backgrounds. These methods have made notable strides in improving detection accuracy, but they do not fundamentally address the challenges of detecting small objects and mitigating the interference from complex background information in remote sensing images23. Therefore, there is still room for further optimization in existing methods to overcome these two major difficulties. The current research contributes by integrating shape-specific bounding boxes and adaptive kernels, outperforming prior works in handling small objects and backgrounds. Small object detection in remote sensing images Small object detection in remote sensing images has long been challenging. In recent years, researchers have proposed various methods to improve the accuracy and robustness of small object detection24–26. The introduction of the YOLO-based algorithm significantly enhanced the capability of small object detection and real-time performance, leading to improved algorithms designed specifically for small object and multi-scale target detection. These improvements have been widely applied in various remote sensing target detection scenarios27–30. Subsequent advancements have also emphasized enhancing the precision of models in detecting small-scale targets31, including mini-YOLO variants that optimize for lightweight detection in constrained environments32,33. The above YOLO-based object detection methods focus on improving small object detection performance through multi-scale feature learning. However, few scholars have attempted to address challenges such as background interference, dense object distribution, missing contextual information, and the sensitivity of the regression branch. To tackle the issue of complex background interference in remote sensing images, CRNet was proposed18, which replaces traditional bounding boxes with circular bounding boxes to reduce background interference in remote sensing imagery. However, the present study primarily emphasizes experimental validation of the method’s effectiveness, while the theoretical principles underlying this approach remain insufficiently explored, which also constitutes one of the key research emphases of this paper. Dynamic adjustment of receptive field In remote sensing image object detection, background information is not always irrelevant, the surrounding image context may carry important information to enhance target detection accuracy34. This issue has received significant attention from researchers. The Receptive Field Expansion Block35 was designed to capture contextual features surrounding each object and propagate them to the feature maps at all levels of the feature pyramid. Additionally, improvements to RFB include using Kronecker convolution for feature extraction, which enhances the global semantic and local contextual information on the feature map by expanding the coverage of individual pixels over the original image36. However, extending the receptive field inevitably introduces interference from background information. Multi-scale convolutions aggregate features at different kernel sizes for richer representations but lack dynamic selection; dilated convolutions expand fields without increasing parameters but can cause gridding artifacts. In contrast, the LSK Network (LSKNet) dynamically adjusts the spatial receptive field based on the target’s size, allowing for more efficient utilization of background information in remote sensing images, thereby improving detection accuracy37. Empirical comparisons in our main experimental results (Sect. 5.3) show LSK outperforms fixed multi-scale (YOLO-v10) or dilated approaches (RetinaNet) in precision for our dataset, justifying its selection. Although the above methods have achieved remarkable results for a specific problem, they have overlooked key challenges in small object detection—multi-scale complexity, background interference, dense targets, context absence, and regression instability. Problem statement Missed storage tank detection caused by sensitivity in the regression branch Fixed IoU thresholds in storage tank detection amplifies regression sensitivity for small objects. While IoU evaluates prediction accuracy during training, its inherent geometric sensitivity causes drastic fluctuations for minor positional shifts (Fig. 1). These instabilities lead to loss function misguidance and subsequent exclusion of small storage tanks during model optimization, resulting in false negatives. IoU variations for bounding boxes of different sizes with the same offset. False target detection caused by improper receptive field setting Fixed receptive fields in object detection models amplify false detections for storage tanks. While expanding the field aids target verification through background context (Fig. 2), excessive enlargement introduces noise. This forces the model to misinterpret ambiguous structures as storage tanks. A dynamic receptive field selection mechanism is thus critical to balance context utilization and noise suppression. Comparison of different receptive field sizes. Methodology This section presents two methods for addressing the two key issues mentioned in Sect. 3. Circular bounding boxes (Sect. 4.1–4.4) solve the problem of missed storage tank detection due to regression branch sensitivity. The LSK module solves the problem of false target detection due to an improper receptive field setting (Sect. 4.5). The proposed method combines these by modifying the YOLO-v10 framework: circular bounding boxes replace rectangular ones in data processing, anchors, and loss functions, while LSK integrates into the backbone for adaptive feature extraction, enabling end-to-end optimization for circular targets in complex scenes. We selected YOLO-v10 as the base due to its state-of-the-art real-time performance and NMS-free design. Circular bounding box Although general object detection algorithms are designed to identify various objects in multiple scenarios, such as autonomous driving, security monitoring, and retail, tank detection presents unique challenges and opportunities. It is mainly used to identify cylindrical tanks in aerial or satellite images, which appear as circles when viewed from above. This contrasts with general object detection, which processes objects of arbitrary shapes and sizes; therefore, tank detection datasets consist primarily of circular targets with consistent shapes, varying sizes, and complex backgrounds. To adapt to these characteristics, our method uses circular bounding boxes, which better match the geometric properties of storage tanks than the rectangular bounding boxes commonly used in general object detection. This design better matches the shape of the targets and enhances detection performance by reducing background noise and improving localization accuracy. Storage tanks typically appear circular in overhead remote sensing images. Compared to rectangular bounding boxes, circular bounding boxes (formed by inscribed circles of rectangles) reduce redundant background area (Fig. 3) and mitigate IoU instability for small objects. Comparison of the enclosed area between rectangular and circular bounding boxes. The regression sensitivity of small storage tanks is amplified by significant IoU fluctuations under positional shifts. To validate circular bounding boxes’ superiority, we prove their IoU fluctuation is smaller than rectangular bounding boxes’ under minor displacements. Let’s assume a rectangular box with side lengths of 2a and 2b, and its inscribed circle with a radius r. We first analyze the IoU of the rectangular box. Let the center points of the rectangular and circular bounding boxes be C1, and the center point after a slight displacement be C2, with a d displacement. The intersection area of the two rectangular bounding boxes can be approximated by Eq. (1). Thus, the union area of the two rectangular bounding boxes is given by For a displacement , the IoU of the rectangular bounding boxes can be approximated by Next, we analyze the IoU of the circular bounding boxes when , the intersection area of the two circular bounding boxes can be approximated by The union area of the two circular bounding boxes is given by Thus, the IoU of the two circular bounding boxes is given by Since the radius of the inscribed circle in the rectangle is , i.e., and . In Eq. (3), represents a sum of first-order terms in d, while in Eq. (6), is a second-order term. It is evident that for , the change in the first-order terms is more significant than that of the second-order terms. Thus, the IoU fluctuation caused by slight shifts in rectangular bounding boxes is greater than that of their inscribed circles. Next, we conduct numerical simulation experiments to validate the conclusion above. We configure two sets of overlapping rectangles with side lengths of 1 and their corresponding inscribed circles, simulating the positional displacement of both rectangular and circular bounding boxes. One set of rectangles and their inscribed circles are moved simultaneously from the initial position along the rectangle’s diagonal, as shown in Fig. 4. Illustration of the numerical experiment setting. Figure 5a shows how the IoU of the rectangular and circular bounding boxes changes with the increase in displacement. It can be observed that, under the same movement speed, the IoU change of the circular box is always smaller than that of the rectangular box, with a crossover point occurring only when the displacement approaches half the side length of the rectangular box. Figure 5b shows the rate of IoU change for the rectangular and circular bounding boxes as the displacement increases. The results show that, for slight displacements, the IoU change rate of the circular box is smaller than that of the rectangular box. IoU variation patterns of rectangular and circular bounding boxes under slight displacement. (a) IoU variation. (b) IoU rate of change. The experimental results indicate that, compared to the rectangular box, the circular box of the same size exhibits significantly smaller IoU variation and rate of change under slight displacement, which is consistent with the theoretical findings. Furthermore, to assess the reliability of the simulation results, a comparison was made of the average IoU and uncertainty range (± 1 standard deviation) of rectangular bounding boxes and circular bounding boxes as the offset increased from 0 to 0.5. To simulate real-world position inaccuracies, random noise of ± 0.5 was applied to the x and y offsets. The experiment was repeated 10,000 times to obtain statistically reliable measures of the average IoU and standard deviation for each bounding box under each offset. as illustrated in Fig. 6. Uncertainty range of rectangular bounding boxes and circular bounding boxes. As illustrated in Fig. 6, the utilization of circular bounding boxes results in lower average IoU and more constrained uncertainty ranges, thereby ensuring more consistent IoU values in the presence of noise. This finding suggests that the symmetrical geometric configuration of circular bounding boxes may serve as a mitigating factor against the impact of positional perturbations. In contrast, the linear overlap of rectangular boxes appears to amplify noise effects. Data preprocessing We first adjusted the data processing section to accommodate the detection requirements of circular bounding boxes. In the original YOLO-v10, the target annotations typically include the class label, the center coordinates of the rectangular box, the width, and the height of the box (x, y, w, h). However, for circular bounding box detection, the annotation information is modified to include the class label, the center coordinates, and the radius (x, y, r). Therefore, we need to calculate the center coordinates and radius of the inscribed circle within the rectangular bounding box, in order to convert the dataset into circular bounding box annotations. Specifically, the x and y coordinates remain unchanged, and the radius r is taken as half of the shorter side of the rectangle. To ensure the quality of circular bounding box annotations, we validated the conversion process from rectangular to circular annotations. The radius was calculated as half the shorter side of the rectangular box, and center coordinates were retained. To address potential labeling bias, particularly for irregular or partially occluded tanks, we conducted a manual review of the dataset to verify the accuracy of the inscribed circle annotations. For partially occluded tanks, where the circular shape may be incomplete, we used a semi-automated tool to adjust the center and radius based on visible tank boundaries, ensuring alignment with the true circular shape. In cases of severe occlusion (less than 50% of the tank visible), annotations were flagged for exclusion to avoid introducing bias. This validation process minimized errors but may still face challenges with highly irregular tanks, which are less common in our dataset but warrant further investigation in future work. Still, with these adjustments, all four-dimensional data processing flows involving (x, y, w, h) in the model optimization and detection framework were adapted to the three-dimensional (x, y, r) format. Additionally, modifications were made to the anchor box structure and the loss function to accommodate circular bounding boxes and the new data dimension. Anchor box structure Although YOLO-v10 is a non-maximum suppression (NMS)-free object detection method, it still employs an initial anchor box structure to facilitate model optimization, helping the model better predict the location and size of objects during training. Since the detection boxes have been adjusted to circular shapes, the design of the anchor box structure also needs to shift from rectangular to circular. The first step is to select an appropriate radius for the circular anchor boxes. Let represents the center coordinates of the circle, and is the radius of the circle (unit: pixels). Each circular anchor box can be represented by Eq. (7). The purpose of using K-Means clustering is to determine representative anchor box radii that cover varying target sizes efficiently, optimizing for multi-scale detection. Like the clustering of rectangular box widths and heights in the YOLO model, multiple anchor box radii are required to cover detection targets of varying sizes. We apply the K-Means clustering algorithm to the ground-truth radii in the training dataset to determine a representative set of anchor box radii . Specifically, to effectively cover most target sizes, we initially selected 9 anchor boxes as the number of clusters K, following the empirical approach used in the original YOLO model. To further optimize the number of clusters, we applied the elbow method to determine the most appropriate K. By calculating the clustering error for different numbers of clusters, we selected the number of clusters at the inflection point of the loss curve as the final K value. After clustering, the center values of each cluster represent the corresponding anchor box radii. These radii are used as anchor box scales for subsequent target detection model training. At the same time, to ensure a uniform distribution of anchor boxes on the feature map, each point on the feature map is associated with multiple anchor boxes, each corresponding to a different radius. Let represent a point on the feature map, and denote the radius of the j-th anchor box. The anchor box assignment on the feature map is expressed as This strategy ensures that objects of different sizes can be detected across multiple feature scales. During training, the process of matching anchor boxes to ground truth bounding boxes is adjusted based on the distance between the predicted and actual radii and the center coordinates. When the IoU between the predicted circle and the ground truth circle exceeds a predefined threshold, as expressed in Eq. (9), the anchor box is considered to match the ground truth bounding box. CircleLoss loss function In the model optimization process, traditional YOLO models typically use loss functions based on rectangular bounding boxes, including bounding box localization loss (e.g., Mean Squared Error (MSE) loss), IoU loss, and object classification loss. However, the geometric properties of circular bounding boxes differ from those of rectangular bounding boxes, necessitating a redesign of the loss function to accommodate the needs of circular object detection. Therefore, for circular object detection, we introduce a custom CircleLoss function. Since the primary focus of this study is storage tanks, which involve single-class object detection, the loss function consists of losses for the center coordinates, radius, and IoU without including a classification loss. The center coordinate loss is crucial for the localization of circular bounding boxes. We use MSE to measure the difference between the predicted and ground truth center coordinates. Let N represent the number of targets, denote the ground truth center coordinates of the i-th target, and denote the predicted center coordinates of the i-th target. The center coordinate loss is defined as . The radius loss is related to the size of the circular bounding box. To ensure that the difference between the predicted and actual radius is minimized, we also use MSE to measure the radius loss. Let represent the ground truth radius of the i-th target, and represent the predicted radius of the i-th target. The radius loss is defined as . The traditional IoU loss is based on the intersection and union areas of rectangular bounding boxes. For circular bounding boxes, the IoU calculation must be redefined. Let represent the intersection area of two circular bounding boxes, and represent the union area of the two circular bounding boxes, the circular IoU can be defined as In summary, IoU loss is defined as . Finally, , and are introduced as weight coefficients to balance the contributions of different loss terms. Combined with the loss of center coordinates, radius loss, and IoU loss, the final loss function is . In this paper, weights , , balance terms, tuned empirically. Large selective kernel In this paper, we combine the LSK with the circular boundary frame detection framework to further improve the performance of tank target detection in remote sensing images. The circular bounding box has the advantage of reducing the interference of redundant background information. Still, its effect depends on the proper use of background information and the ability to suppress it. By incorporating LSK, the size of the convolution kernel can be selected adaptively, and the utilization efficiency of background information can be further optimized based on a circular boundary box. LSK originated from LSKNet, and its core idea is to combine the multi-scale convolution kernel with the selection mechanism to enhance the flexibility and robustness of the network. LSK constructs large convolutional kernels by decomposing them into a series of depth-wise convolutions with progressively increasing kernel sizes and dilation rates. Specifically, the receptive field size of the i-th convolution is represented as Eq. (15). The kernel size ki and dilation rate di of the i-th depth-wise convolution can be defined as After obtaining features with receptive fields at different scales, the LSK module dynamically selects the most relevant spatial context areas for different targets using a spatial selection mechanism. This mechanism first extracts spatial feature descriptors through average pooling and max pooling, and then uses convolutional layers to transform these descriptors into spatial attention maps. Each spatial attention map is then converted into a spatial selection mask through a sigmoid activation function, which is used to weight the features corresponding to the decomposed large convolution kernel as Eq. (17). This dynamic adjustment allows smaller kernels for small targets to preserve details and larger ones for big targets to capture context. Through the above mechanism, the LSK module can adaptively adjust the network’s receptive field to better capture the contextual information required by different targets. In complex backgrounds, this dynamic adjustment capability enables the network to reduce background interference, improve detection accuracy for small objects, and, at the same time, effectively leverage more contextual information for large targets, thereby enhancing overall object detection performance. We replaced the C2f Block in the YOLO-v10 backbone network with the Large selective Kernel Block (LSKBlock) layer. The overall architecture of the modified model is shown in Fig. 7. Overall model architecture diagram. (Embedded satellite image examples were sourced from the composite dataset described in Sect. 5.1: DIOR38, NWPU_RESISC4539, NWPU VHR-1040, TGRS-HRRSD41, and self-built dataset42. In the LSKBlock, the network dynamically selects the most appropriate convolution kernel by combining convolution kernels of different scales (e.g., ) to accommodate the multi-scale nature of the object in the input image. If the input feature graph is X, is the i-th convolution kernel, and is the corresponding selection weight and satisfies , the output of the selective convolution layer Y is In remote sensing images, detecting small objects usually requires a smaller convolution kernel to preserve detailed features. In comparison, large objects or objects requiring more extensive background information may need a larger convolution kernel to capture the global context, which is achieved by the selection function of Eq. (19). The function dynamically selects the appropriate convolution kernel output based on the statistics of the feature graph and the size of the target. Integrating LSK, YOLO-v10 can effectively combine the advantages of LSKNet to enable it to dynamically adjust the receptive field according to the size of the detection target and flexibly select the utilization rate of remote sensing image background information when processing the multi-scale storage tank target detection task in remote sensing images, thus reducing the occurrence of false detection. Experiments Datasets This experiment selects the data specially used for storage tank target detection, composed of 5 remote sensing datasets. Including the DIOR dataset38, NWPUU_RESISC45 dataset39, and NWPU VHR10 dataset40, TGRS-HRRSD dataset41, and a self-built dataset42. Other irrelevant targets in these data sets were excluded, and only images and label files of different storage tank targets under various complex backgrounds were retained. The basic information of the dataset is shown in Table 1. Table 1Basic information about the datasets.DatasetResolutionImage size/pixelNumber of images containing storage tanksNumber ofstorage tanksDIOR0.5–30 m800*800124420,361NWPUU_RESISC450.2–30 m256*25668812,405NWPU-VHR-100.5–2 m(500–1100)*(500–1000)1651698TGRS-HRRSD0.15–1.2 m(152-10569)*(152-10569)8974406Self-built dataset0.5–3 m512*5125747205Total————356846,075 Basic information about the datasets. (500–1100)* (500–1000) (152-10569)* (152-10569) Since the object detection frame is adjusted to a circular bounding box in this paper, the labels of the dataset need to be adjusted accordingly. According to existing label positions in the dataset, the rectangular box is adjusted to the corresponding tangent circle as the circular boundary box, and the label information is modified to include the category label, the circle center coordinate, and the radius of the three-dimensional label. Figure 8 shows the remote sensing image after the label visualization. The remote sensing image after the label visualization. (Embedded satellite image examples were sourced from the composite dataset described in Sect. 5.1: DIOR38, NWPU_RESISC4539, NWPU VHR-1040, TGRS-HRRSD41, and self-built dataset42. Circular bounding box labels were visualized using Matplotlib version 3.8.0 (https://matplotlib.org/) in Python 3.9.0 and Labelme version 1.3.1(https://github.com/wkentaro/labelme)) Finally, we divided the dataset into training sets, validation sets, and test sets according to the ratio of 7:2:1, in which the training set had 2498 images, the validation set had 714 images, and the test set had 356 images. Experimental environment and setting The experiment was run in the following hardware environment: Intel Xeon(R) Gold 6226R@2.90 GHz 2.89 GHz CPU, 256GB RAM, NVIDIA Quadro RTX 6000 GPU, Windows 10 Professional Edition, CUDA 11.1 and Pytorch 1.10.1. Training uses a 0.01 initial learning rate, 0.01 multiplier, 3-epoch warm-up, 0.937 momentum, and early stopping after 100 epochs without improvement. The evaluation indexes in this paper are Precision (P), Recall Rate (R), and Mean Average Precision (mAP). P represents the probability that the model predicts correctly, and R represents the proportion of the sample the model predicts in the real sample. mAP is a more comprehensive evaluation index in the target detection task, which can evaluate the detection ability of the model more comprehensively. We use mAP@0.5 to evaluate the model’s performance, representing the mean accuracy at an IoU threshold 0.5. Main experimental results In the model training stage, we adopted the dynamically adjusting the learning rate to accelerate the training convergence and avoid falling into the local optimal solution. Specifically, we initialized the learning rate at 0.01 to facilitate rapid learning at the start. To ensure more fine-tuned convergence, using a learning rate multiplier at 0.01 to gradually reduce the learning rate during training. We also employed a warm-up period of 3 epochs to stabilize the training in the early stages. During this warm-up, the learning rate started from a low value and gradually increased to the initial value of 0.01. Additionally, the momentum was initialized at 0.937 to help accelerate gradient descent and avoid oscillations in the optimization process. To prevent the model from overfitting in the later stage, we adopted the early stop method during the training stage. When there was no performance improvement of the model within 100 consecutive epochs, the training was stopped, and the model with the best performance was taken as the final result. Finally, the model reached the early stop condition after 153 training epochs. The changes in its indicators are shown in Fig. 9. In the early stages, the validation loss fluctuates significantly compared to the training loss, but it gradually stabilizes as the number of epochs increases. During the initial training stages, factors such as overfitting, differences in data distribution, learning rate settings, and gradient noise cause the model to perform inconsistently on the validation set. As the number of epochs increases, however, the model learns more generalizable features, the optimization process converges, and the model adapts to the distribution of the validation set. This results in a gradual decrease in the fluctuation of the validation loss and eventual stabilization. Loss change curve during model training. We comprehensively consider the precision and recall of the validation set and select the best round model with the corresponding epochs of 85, precision of 0.911, recall of 0.902, and mAP@0.5 Is 0.931. From the loss curve in Fig. 9, it can also be seen that after 85 epochs of the model, although the training loss is still declining, it is verified that the loss has converged and no longer has an obvious downward trend. If the training is continued, the model will be overfitted on the training set, reducing the model’s generalization ability and affecting the final detection effect. Therefore, we saved the 85th-round model and used it for the final test. To show the impact of the method proposed in this paper, we selected some representative test results and showed them in Fig. 10. Figure 10a shows a batch of test images after tag visualization; Fig. 10b shows the actual detection results of the batch of images, with the confidence of each detection target. Comparison of detection results and labels in the test set. (a) Original labeled image. (b) Actual test result. (Embedded satellite image examples were sourced from the composite dataset described in Sect. 5.1: DIOR38, NWPU_RESISC4539, NWPU VHR-1040, TGRS-HRRSD41, and self-built dataset42. Circular bounding box labels were visualized using Matplotlib version 3.8.0 (https://matplotlib.org/) in Python 3.9.0 and Labelme version 1.3.1(https://github.com/wkentaro/labelme)) The results in Fig. 10 include both regular-sized targets and small objects, along with complex background objects. The evaluation metrics and overall detection performance are satisfactory. Notably, the trained model still detected some targets located at the edges of the image, which were not fully labeled in the validation set due to their incomplete appearance. These detection results, however, do not positively impact the evaluation metrics, indicating that the model’s actual performance may slightly exceed what the metrics reflect. Of course, our proposed method does not guarantee 100% accuracy. The model may still produce false positives when confronted with objects that are difficult to distinguish from storage tanks. For example, in Fig. 10, the model misidentified a nearly circular stadium as a storage tank. We speculate that the possible reason is that most storage tanks in remote sensing images are light-colored, such as white, while there are fewer dark-colored tank samples in the training set. As a result, the model may lack sufficient training data for dark circular objects, leading to misidentifications. To further verify the effectiveness of the proposed method, we selected five representative object detection methods for comparison, including three classic methods: Faster R-CNN43, SSD44, RetinaNet45, and three newer methods: MADnet46, YOLO-v7 + CBAM8, YOLO-v1019. Since these methods all use traditional rectangular bounding box detection frameworks, the dataset labels used in this study are not directly applicable. Therefore, in the comparative experiments, all methods that do not involve circular bounding boxes still use traditional rectangular label files. To ensure fairness in the experiments, we compared the best-performing model from each method with the results shown in Table 2. Table 2Comparison of results with different methods.MethodsPrecisionRecallmAP@0.5RetinaNet0.8010.7620.630Faster RCNN0.8170.7230.846SSD0.8690.8670.894MADnet0.8390.8100.790YOLO-v7 + CBAM0.8900.8700.900YOLO-v100.8930.8780.915Ours0.9110.9020.931 Comparison of results with different methods. The experimental results indicate that, when trained on the same storage tank detection dataset, our method outperforms others in Precision, Recall, and mAP@0.5, demonstrating that our approach achieves the best performance for storage tank detection in remote sensing images. Model complexity experiment To assess the computational complexity of the constructed model, a comparative analysis was conducted between the improved model and the original YOLO-v10 model. The analysis encompassed four distinct aspects: the total number of layers, the number of parameters, the computational complexity, and the gradient parameters. Table 3 shows the complexity comparison results of the two models. The improved model shows reduced parameters and GFLOPs compared to the original, indicating better efficiency despite added layers, making it more suitable for resource-constrained deployments. Table 3Model complexity experiment comparison results.ModelsNumber of layersNumber of parametersComputational complexityGradient parametersFrames per secondYOLO-v103852,707,4308.4 GFLOPs2,707,41474Ours4712,619,3827.6 GFLOPs2,619,36678 Model complexity experiment comparison results. The experimental results in Table 3 demonstrate that the number of layers in the improved model increased from 385 to 471, representing an approximate 22% increase. This finding suggests the presence of a more intricate network architecture, one that is adept at capturing more complex features. This is of particular importance for tasks involving tank detection, as tanks in remote sensing images are often small and have complex backgrounds. A more intricate network has been shown to enhance the extraction of features, thereby improving the accuracy of detection. However, despite the increase in the number of layers, the number of parameters in the improved model decreased by 3.2% (from 2,707,430 to 2,619,382), computational complexity decreased by 9.5% (from 8.4 GFLOPS to 7.6 GFLOPS), and the number of gradient parameters also decreased by 3.2% (from 2,707,414 to 2,619,366). This phenomenon is attributed to the presence of more sublayers within the LSK module, each of which has been carefully optimized. Notwithstanding the augmented number of layers, the majority of these layers are deep convolutions with a paucity of parameters. These enhancements suggest that the enhanced model exhibits enhanced efficiency in parameter utilization, reduced computational demands, and optimized memory usage during training. To evaluate the real-time performance of the proposed model, we measured inference latency and frames per second (FPS) on the test set using the experimental environment described in Sect. 5.2. The improved model achieved an average inference time of 12.8 ms per image, corresponding to approximately 78 FPS, compared to 13.5 ms per image (74 FPS) for the original YOLO-v10 model. This improvement aligns with the 9.5% reduction in computational complexity, demonstrating that the proposed method maintains real-time performance suitable for large-scale remote sensing applications. These results indicate that the model is efficient for deployment in resource-constrained environments, such as on-site methane emission monitoring systems. This characteristic is of significant importance in engineering applications. The processes of tank detection and methane emission monitoring typically entail the analysis of large-scale remote sensing images, a task that demands significant computational resources. The enhanced model exhibits a reduction in computational complexity and parameter counts, consequently leading to a decrease in deployment costs and energy consumption. Moreover, it demonstrates an increased suitability for operation on devices with limited resources, while maintaining or enhancing detection performance. Ablation experiment We designed a set of ablation experiments to validate the contribution of each improvement in our proposed method to the accuracy of storage tank detection in remote sensing images. Compared to the standard YOLO-v10 method, our approach introduces two main enhancements. First, we modified the detection framework to use circular bounding boxes. Second, integrating an LSK into the backbone network. While designing the ablation experiments, we aimed to isolate the contributions of circular bounding boxes and the Large Selective Kernel (LSK) within the YOLO-v10 framework, recognizing their dependency on this baseline for effective operation. These modules, as incremental enhancements, rely on the framework’s infrastructure and cannot be trained independently, which guided our decision to use YOLO-v10 as the reference point. The experimental design and results are shown in Table 4. Table 4Ablation experiment settings and results.Experiment settingsPrecisionRecallmAP@0.5Computational complexityGradient parametersYOLO-v100.8930.8780.9158.4 GFLOPs2,707,414YOLO-v10 + Circular bounding box0.8950.8900.9198.0 GFLOPs2,682,390YOLO-v10 + LSK0.9010.8800.9238.2 GFLOPs2,671,820Ours (YOLO-v10 + Circular bounding box + LSK)0.9110.9020.9317.6 GFLOPs2,619,366 Ablation experiment settings and results. The experimental results show that our proposed method achieved the best performance in Precision, Recall, and mAP@0.5 for the storage tank detection task in remote sensing images when both circular bounding boxes and the LSK were introduced. This demonstrates that the two improvements complement each other, collectively enhancing the detection accuracy of storage tanks. When each improvement was individually incorporated into YOLO-v10, both positively impacted the experimental results. When circular bounding boxes were integrated into YOLO-v10, the IoU fluctuations for small objects were reduced, leading to the detection of more small objects and a reduction in missed detections, significantly improving Recall. Precision and mAP@0.5 also showed slight improvements, indicating that circular bounding boxes are well-suited for detecting near-circular objects like storage tanks. On the other hand, incorporating LSK into the backbone network dynamically adjusted the receptive field, strengthening the model’s detection capability in complex environments and reducing false positives, which resulted in a notable improvement in Precision. Meanwhile, Recall and mAP@0.5 also displayed minor increases, suggesting that the LSK enhances the model’s ability to detect targets of varying scales by dynamically adjusting the receptive field. As illustrated in Table 4, a trade-off between cost and performance is evident. For instance, increasing LSK alone or switching to circular bounding box detection reduces the complexity of the original YOLO-V10 (lower parameters/GFLOPs), yet the overall accuracy improvement is not significant. Conversely, full-combination optimization has been demonstrated to enhance overall efficiency. These experimental results demonstrate that our method effectively combines the strengths of YOLO-v10, circular bounding boxes, and the LSK, significantly improving the detection performance of storage tanks in remote sensing images. The method enhances precision and recall, enabling the model to more accurately identify storage tanks and effectively reduce false positives for non-tank targets. Discussion The experimental results demonstrate that the combination of circular bounding boxes and LSK achieves the optimal performance of the proposed method in the storage tank detection task (accuracy: 0.911; recall rate: 0.902; mAP@0.5: 0.931). Compared to baselines like YOLO-v10 (mAP@0.5: 0.915), our method improves by 1.8%, primarily due to better handling of small objects (via circular boxes) and backgrounds (via LSK), unlike Faster R-CNN (0.846) which suffers from slower speed and regression issues, or SSD (0.894) limited by fixed anchors. The distinction arises from our shape-specific design and adaptive kernels, tailored to circular tanks, whereas others are general-purpose and less efficient for this domain. The physical and engineering significance of these improvements can be further understood from the following aspects: (1) The function of circular bounding boxes is as follows: The incorporation of circular bounding boxes into the YOLO-v10 model led to an enhancement in the recall rate from 0.878 to 0.890, signifying a substantial improvement in the detection capability of diminutive storage tanks. From a physical perspective, this enhancement can be attributed to the high degree of congruence between circular bounding boxes and the geometry of storage tanks, which minimizes the impact of superfluous background pixels in conventional rectangular bounding boxes. Theoretical analysis (Sect. 4.1) demonstrates that circular bounding boxes demonstrate reduced variability in IoU calculations for small objects. From an engineering perspective, this translates to diminished sensitivity to positional deviations, consequently leading to a reduction in false negatives. This characteristic is of particular importance for small storage tanks in remote sensing images, as these targets typically occupy only a small number of pixels and are easily missed by traditional methods due to regression instability. Regarding generalizability, while circular boxes reduce background coverage, the model still learns from diverse training data including partial backgrounds around circles, allowing robustness to noise as validated in simulations (Fig. 6). (2) The role of LSK is twofold: first, its incorporation into YOLO-v10 enhances the accuracy from 0.893 to 0.901, thereby demonstrating its efficacy in reducing false detections. From an engineering perspective, LSK adapts to the multi-scale characteristics of storage tanks by dynamically adjusting the receptive field. The capacity of small storage tanks to preserve details is contingent upon a smaller receptive field, while the capacity of large storage tanks or targets in complex backgrounds to obtain contextual information is enhanced by a larger receptive field. This adaptability corresponds physically to the model’s selective utilization of background information, enabling it to better distinguish storage tanks from similar circular objects (e.g., the stadium in Fig. 8), thereby enhancing detection accuracy. (3) The following section will explore the synergistic effects of the two methods. The employment of circular bounding boxes in conjunction with LSK has been demonstrated to enhance the model’s performance, as evidenced by the enhancement in accuracy and recall rates to 0.911 and 0.902, respectively. This synergistic effect indicates that the two methods complement each other in addressing the challenges of storage tank detection (e.g., small object misses and false positives in complex backgrounds). The utilization of circular bounding boxes provides prior knowledge regarding the shape, while LSK enhances the model’s adaptability to targets and backgrounds of various scales. This combination offers an efficient engineering solution for methane emission monitoring, achieving a balance between high coverage (high recall) and high reliability (high accuracy). (4) To better understand the limitations of our proposed method, we analyzed common failure cases observed in the test results, as shown in Fig. 10. A notable misdetection example is the misclassification of a nearly circular stadium as a storage tank, primarily due to its similar shape and high contrast in remote sensing imagery. This error likely stems from the underrepresentation of dark-colored circular objects in the training dataset, as most storage tanks are light-colored (e.g., white or metallic). Other failure cases include missed detections of partially occluded tanks, where shadows or overlapping structures obscure key features, and false positives in densely clustered scenes, where small circular objects are mistaken for tanks. To mitigate these issues, future improvements could involve augmenting the dataset with more diverse samples, including dark-colored circular objects and occluded tanks, to enhance model robustness. Additionally, incorporating color and texture features into the LSK module could help distinguish tanks from similar structures, reducing false positives. These strategies aim to address the specific challenges of complex backgrounds and morphological similarities in remote sensing imagery. The significance of these improvements extends beyond mere metric improvements. The proposed methodology directly addresses two fundamental challenges in the field of storage tank detection in remote sensing images. Firstly, it considers the regression sensitivity of small objects, and secondly, it explores the interference from complex backgrounds. By enhancing detection performance, this method provides the oil and gas industry with a more precise tool for locating methane emission sources, thereby supporting the implementation of environmental monitoring and climate change response strategies. Potential limitations include limited generalization to non-circular objects, as the circular design is tailored to tanks; future directions could involve hybrid bounding boxes for diverse shapes and further LSK optimizations for real-time applications. Conclusion This paper presents a method for storage tank detection from remote sensing images based on circular bounding boxes combined with LSK, aiming to tackle the problems of missed mall objects and false detections in complex backgrounds commonly encountered in traditional detection methods. Circular bounding boxes are introduced to reduce missed detections of small objects, which mitigate the issue of missed small objects due to regression branch sensitivity as shown in complete theoretical proof provided in this study. LSK is employed to dynamically adjust the receptive field to address false positives in complex backgrounds, enabling selective use of background context information from remote sensing images, thereby enhancing detection accuracy. Experimental results show that, in the specific application scenario of storage tank detection in remote sensing images, the proposed method significantly improves precision, recall, and mAP@0.5 compared to both classical and latest object detection methods, achieving scores of 0.911, 0.902, and 0.931, respectively. This method addresses two key challenges in storage tank target detection: sensitivity to small objects and complex background interference. The technology under discussion offers an efficient solution for storage tank localization in the petroleum and natural gas industry, facilitating the monitoring and tracking of greenhouse gas emissions. This, in turn, supports efforts to mitigate climate change. In the future, this framework is expected to be extended to detect more methane emission facilities, and plans are underway to expand the dataset to enhance the model’s ability to distinguish between storage tanks and similar background objects. Author contributions Y.L. proposed a research plan and conducted experimental design, and was the main contributor to manuscript writing. Y. W. and W. H. revised the manuscript. Z. Z. prepared the datasets. All authors reviewed the manuscript. Funding This work was supported by China Petroleum and Chemical Corporation under the project [Research on Integrated Ground and Atmospheric Monitoring, Tracing, and Accounting System for Methane and VOCs in Oilfields]. Award Number: 322121. Data availability The datasets generated or analyzed during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated or analyzed during the current study are available from the corresponding author on reasonable request.