Enhancing highway security and wildlife safety: Mitigating wildlife–vehicle collisions with deep learning and drone technology

2.1 Tourist security and safety

Tourism is a competitive business, with a core objective of attracting tourists and consequently influencing their tourism destinations. A tourism destination is a specific location that contains attractions that improve visitors’ travel experiences [26]. In South Africa, the competitive advantage of the tourist sector is popularly backed by science, technology, information, and innovation. The conservationists in this country focus on managing natural resources by creating innovative solutions that contribute to fostering tourist destinations. In this regard, South Africa has a well-connected network of 6.3% national parks and private nature reserves [27]. The latter trend is great because it meets the demands of environmentally sensitive visitors. Even though a conducive environment is implemented for tourists in most of these parks, several factors limit the tourism industry’s effectiveness in South Africa, including (1) fewer funds allocated to the tourism sector; (2) limited merging of local communities into tourism; (3) inadequate education, training, and awareness concerning tourism; and (4) tourism security and safety. These challenges discourage potential tourists from visiting South Africa, negatively reducing tourism demand and leading to low tourist destinations. To minimize the challenges, the Western Cape government of South Africa implemented a six-leveler strategy for tourism safety, and these are: (1) a tourism safety forum, which forms a formally structured, multi-stakeholder, and collaborative safety forum; (2) establishing a multi-stakeholder enforcement task team; (3) provision of tourists safety response by providing assistance and supporting tourists while in distress; (4) to identify and develop tourists safety and security technology platform that can act as an enabler to help address and solve core tourists safety risks; (5) to develop a tourism safety comprehensive communication strategy that integrates all important stakeholders; (6) tourism safety interventions and partnerships by identifying synergies that enable funding of cross-platform initiatives and skills developments. Our work implements strategy four of the Western Cape government, one of the provinces of South Africa, which proposes to innovate a technology platform that helps address tourists’ safety and security risks [28]. We note that tourism security is a major constraint to overseas tourism growth. Safety and security considerations are important aspects of increasing tourism destinations in a country. The success or failure of the destination of tourists depends on the possibility of providing a safe and secure environment for visitors [29]. This study seeks to improve the safety and security of tourists on highways by intervening using artificial intelligence algorithms to minimize wildlife encounters with tourists. Tourists on South African highways need to be made aware of the presence of wildlife fences because some visitors tend to get out of their vehicles near demarcated wildlife places to enjoy the scenery with no idea that some places have dangerous wildlife. Argumentatively, the machine learning algorithms developed in this study can be incorporated into maps and deployed on car dashboards or driverless cars to automatically detect a wildlife fence and record and save the GPS coordinates and other parameters. The purpose of this work is to inform tourists of the presence of wildlife and wildlife fencing, providing security by minimizing encounters between wildlife and tourists, which finally leads to the mitigation of WVC.

2.2 Wildlife roadkill area detections

Roadkill areas vary, spatially and temporally [30]. Researchers in developed economies use intelligent animal detection systems (ADSs) [15–17, 31–44] to detect roadkill areas. These systems are usually costly, and unaffordable to African governments, thus, limiting their implementation in Africa. Popularly, in time series analysis, the roadkill areas are detected using latitude and longitude coordinates collected by researchers and highway maintenance personnel. However, these detected roadkill areas tend not to align well with sites of most serious animal crossing attempts due to imperfect detections [45]. To address the weakness of imperfect detections and the failure of African governments and organizations to implement a cheaper and novel approach for mitigating WVC, we propose and design an artificial intelligence mitigating strategy using classification and segmentation techniques to detect wildlife fences and their associated features called insulators in WVC locations. The algorithms can further be deployed on client-edge devices to assist drivers in detecting fences when driving the car. The camera can see the wildlife fence image and save the GPS coordinates. Further, such a dataset can be integrated into maps and GPS satellites to predict and visualize future WVC locations to drivers’ devices far away before they arrive at a WVC location. Such solutions should be ethically aligned to ensure no harm is caused to animals and humans [46].

2.3 WVC prevention

The WVC is a global problem that has escalated in developed and developing countries. United States registers 1,500,000 WVC cases, and Canada experiences the death of 4–8 animals per day [8]. In South Africa in 2015, the greater Mapungubwe transfrontier conservation area record shows 36 and 991 WVC on unpaved and paved roads, respectively [6]. To prevent WVC, mitigation measures are implemented. These are overpasses, underpasses, wildlife fencing, break-the-beam [15,31–34], area cover [15,16,31,35–38], buried cable [17,39,40], driver assistance [41], autonomous vehicles [42], and mobile mapping ADSs [43,44]. Intelligent systems are popular in developed countries and these are not implemented in South Africa.

In this study, we propose and implement a novel mitigation measure that is suitable for the South African economy. It is cheaper and uses existing client-edge devices such as car dashboards and mobile phones. The measures use wildlife fencing, drone technology, classification, and segmentation algorithms to detect wildlife fences and the associated features. The algorithm can be incorporated into maps to detect fences and alert tourists and drivers about wildlife fences on highways. This approach improves the security and safety of tourists and wildlife. In addition, it mitigates WVC.

4.1 Study area

We took the images used in this study at Amakhala and Lalibela Private Game Reserves located in the Eastern Cape province of South Africa (Figure 2). These game reserves are located on the highways on a flat and low terrain surrounded by green vegetation. We captured images of the experimental site using a drone and still cameras. To capture aerial images, a DJI Phantom 4 Pro UAV drone was used, and Nikon Z6 Mk 2 still camera helped in capturing still images. The Nikon Z6 Mk 2 contains a 24.5MP FX-Format BSI CMOS Sensor, and Actual: 25.28 Megapixel, with effective: 24.5 megapixel ( 6,000 × 4,000 ) ISO 100-51200, and contains in-body 5-axis vibration reduction sensor which incorporates a 273-point phase-detection AF system, mechanical focal plane shutter, electronic shutter, with the image file format is RAW. The UAV has a 1 20MP CMOS sensor, with a gimbal-stabilized 4K60 20MP imaging, Ocusync Transmission, flight autonomy with redundant sensors, it operates in five directions of obstacle sensing, and four directions of obstacle avoidance, with a maximum flight time of 30 min, and hovering accuracy ± 0.98 ′ /0.3 m and horizontal with vision positioning ± 0.33 ′ ∕ 0.1  m. In addition, it has a vertical with vision positioning and still image support – DNG. The highest flying height we went with the drone was 204 m.

Figure 2

The Amakhala and Lalibela game reserves, Eastern Cape Province, South Africa, showing sampling points in green. Source: Created by the authors.

4.2 Data acquisition

The data collection comprised three co-authors and one UAV pilot who captured the aerial and still photographs. The photographers captured scenes of wildlife fences based on the research objective. Researchers stated the key features to focus on while photographing the wildlife fences. The data collected contains 52 and 159 images from the drone and still cameras, respectively, as shown in Table 1. For both devices, we stored the photos in .jpeg format. Aerial images contain a height of 5,904 and width of 3,933 pixels. Still, images have a height of 5,904 and width of (3,928 or 3,936) pixels. In both image datasets, we observe very high-resolution pixels.

4.3 Data preprocessing and annotation

The images were in image file format – RAW. This file type is uncompressed and commonly captured by digital cameras and sensors. We processed the images from RAW files to JPEG file format using Adobe Photoshop to convert the format and filter the converted images, totaling 211 from UAV and standalone cameras.

For the classification task, the images were compressed to equal dimensions of 512 pixels by 512 pixels. In both datasets, the images were saved in separate folders. The still dataset contained images captured by the standalone camera; the drone captured aerial images of wildlife fences from the sky. The drone camera images used in this study were 26 from the single fence and 26 from the double fence. We had 159 images from a standalone camera containing 79 single and 80 double fence images (Table 1). In Figure 3, we visualize a random sample of images from our dataset and deploy it on mendeley [50]. Because of a few samples, there was a need to expand the training image size to improve model accuracy. Techniques of data augmentation were implemented in the training dataset sample. The techniques made variations to the training sample, avoiding overfitting, and increasing robustness [51]. Data augmentation transformations applied in this study include horizontal flips, vertical flips, random crops, and a small Gaussian blur of approximately 50% of the total images sample with a random sigma of between 0 and 0.5, strengthening or weakening the contrast in each image. We also added some Gaussian noise per pixel and channel by sampling only 50% of all images; this can change not only the color but also the brightness of the pixels. Further, we brightened and darkened some images by sampling 20% of the images of all cases. Each channel is sampled with the multiplier, which leads to the changed color of the images. Finally, affine transformations are applied to each image by scaling, zooming, translating, moving, rotating, and shearing them. All the augmentations performed were applied randomly. From Table 1, 15 Images (from aerial) and 47 Images (from still) were removed and kept aside as a test set; later, the images were used for testing the model. The remaining images in Table 1 were augmented to generate the training and validation sets in Table 2.

Figure 3

Random samples from wildlife fences datasets for Classification Task [50].

For the segmentation tasks, the original JPEG images have high-resolution pixels as described in Section 4.2. In our observations, the wildlife fence images contain a lot of noise, such as grass and trees, compared to the fence itself. Besides, these images are enormous for machine learning models to process. To minimize the noise, we cropped out noisy parts of the images to 512 × 512 pixels. The sliced images obtained contained both images with no insulators (noise) and images with insulators. We discarded images with only noise and retained images with insulators, as shown in Figure 4, where we input the original images and output the cropped images. For each original image, we reshaped it to obtain an equal dimension image size for easy cropping into smaller images, which were easily processed by the U-Net algorithm. We then sorted the photos into positive and negative regions. Further, we keep the positive regions for further processing and discarded the negative regions.

Figure 4

Split the original image into smaller images of 512 × 512 pixels. Discard images with no insulators and retain images with insulators for further processing [50].

To build the ground truth mask of the insulators, we used the Visual Geometry Group (VGG) Image Annotator (VIA) [52]. The VIA was used to define regions and textual descriptions of the regions. The annotated images were saved in a .json extension file containing the coordinates of the insulators. In this task, each sliced image of a wildlife fence insulator was annotated to label insulators as positive regions. The rest of the objects in the image were labeled as negative regions. The positive regions were identified as 1, and the negative regions were identified as 0, which gave us a gray-scaled image as an output or binary mask. The results from the annotations were image masks for each insulator found on the wildlife fence. Since VIA output had smaller images with a single insulator, to reconstruct a larger image containing all insulators in an image, we concatenate them, as shown in Figure 5 and Algorithm 1.

Figure 5

Input the sliced 512 × 512 pixel image for annotation to generate individual masks for each insulator in the image, later we combined the individual masks to resemble the original image [50].

The generated pairs of 512 × 512 pixel image masks with associated original images, as shown in Figure 5 were fed into a neural network for model training and fitting to learn to find the positive regions (insulators) on the wildlife fence images. The annotated images are shown in Table 3, including 447 and 686 aerial and still images, respectively, extracted from single and double fences. These are split into training, validation, and testing sets. Both datasets contain 80% samples for training, 10% for validation, and 10% for the test.

Figure 6 shows all the processes described above, i.e., from fieldwork for data collection, to the implementation of the recognition and detection models then the testing from field data.

Figure 6

Proposed methodology: The process shows the proposed pipeline from the recognition and detection implementation to the test from field data. Source: Created by the authors.

5.1 CNNs

CNNs have input and output layers, with hidden convolutional layers, pooling layers, and a fully connected layer placed in between. The layers are composed of neurons that contain learnable weights and biases. Each neuron takes in respective inputs, computes a weighted sum by passing it through an activation function, and returns an output [53]. The CNN’s are used to classify images by learning intrinsic image features, clustering these images by similarities, and finally, performing object recognition. A CNN takes in images as input and allocates importance by applying the learnable parameters (weights and biases) to several features in the image, and by doing so, it can differentiate one image from the other [54].

5.2 Transfer learning ResNets

This study leverages transfer learning using the wildlife fences dataset. An approach is known for developing novel applications in image classification, detection, localization, and segmentation tasks. In attempting to improve the performance of CNNs, a transfer learning approach was introduced. Transfer learning technique in machine learning is when we pre-train a model on a given dataset and then re-used it as a building block in a related task [55]. Transfer learning has registered successfully in many applications in machine learning, including image classification [56], image localization and detection [57], as well as image segmentation [58], and has led to the development of CNN models that offer a much greater performance [55,59]. Examples of large transfer learning models include VGG, GoogleNet (Inception), and Residual Network (ResNet). In this study, we adopted ResNet50 architecture because of its power of skip connections that eliminates the vanishing gradient problem. We draw upon these works during model development and tuning hyperparameters when fitting the training data to each binary classification model, as it suggests overfitting is a common issue when transferring learning to validation and testing sets. The ResNet was developed to overcome some challenges experienced when using the initially created transfer learning models. These common challenges include problems with degradation, network optimization, and vanishing or exploding gradient issues. The ResNet50 model uses the concept of shortcut or skipping in the CNN to carry out the task [60,61]. ResNet has shown great performance when applied to computer vision tasks.

5.3 U-Net for segmentation

The architecture of the U-Net network consists of an encoder (downsampler) and decoder (upsampler) with a bottleneck connecting the encoder and decoder [24]. Skip connections join encoder outputs to each stage of the decoder. The decoder or contracting path contains a convolutional network with alternating convolutions of 3 × 3 followed by the ReLU activation unit as shown in equation (1) and max pooling operations as shown in equation (2). Each stage had an increasing number of filters, and the dimensionality of the features was reduced because of the pooling layer. Later, more features were extracted by the bottleneck. Finally, the expansion path or the decoder expanded the output resolution by upsampling the features using 2 × 2 up-convolution back to the original image size as shown in equation (3). To localize the object of interest and features upsampled, the expansive path concatenated them with high-resolution features from the contracting path through skip connections by cropping and concatenating the features. At each upsampling level, we took the output of the corresponding encoder or contracting path and concatenated it before feeding it to the next decoder or expansion path. Two 3 × 3 convolutions and ReLU activation during upsampling were used. The final stage applies an additional convolution of 1 × 1 , which reduced the feature map to obtain the required number of channels and finally produces the segmented image as shown in Figure 7. During the expansion path, the cropping of images makes the image lose contextual information, leading to a network resembling a u-shape. More importantly, it generated contextual information along the network, enabling the network to segment objects in an area by overlapping areas, outputting a pixel-by-pixel mask showing the class of each pixel

where c is the bias vector, w is the weights matrix, a is the input feature map, and b is the output feature map. Here, x is the width, y is the height, k are the features index, and l is the depth of image.

Figure 7

U-net architecture has arrows that shows the direction of the different operations. Feature maps are represented at each layer using the blue boxes. Cropped feature maps are in the light gray boxes. The gray boxes are from the contracting path and dark gray represents the output image channels. Source: Re-created by the authors [24].

5.4 Optimization via meta-learning

We adopted the transfer learning and a meta-learning type of optimization because we had a small dataset. The idea of finding the optimized deep models Ψ ( θ , ⋅ ) that could be efficiently tested on a small dataset using meta-learning is as follows. From the basic optimization problem, we want to minimize a loss function ℒ ( y , Ψ ( θ , X ) ) , which uses the cross-entropy for classification problems

where the vector y with components y i are know expected predictions, and the vector X with entries X i are features network inputs. Assuming a sampled task, T ∼ ρ ( T ) from the distribution of tasks, p ( T ) . Split the entire dataset into training, validation, and testing sub-datasets and randomly choose some N samples of labels from each of the classes and sequentially update the model weights vectors θ i j :

where α is the learning rate, y S i and X S i are the expected input predictions sub-vector and input features sub samples elements of the support set S . Note that θ i 0 ≡ θ . After the update of task-specific weight θ i j , the update of θ now properly follows:

Here, β is the learning rate, y Q i and X Q i are the expected input prediction and feature vector elements of the query set Q . It can be seen that the update of θ is an average of the objective function defined in equation (4) across the query set samples.

6.1 Confusion matrix

The study focused on solving a binary classification problem by evaluating the optimal solution that predicts whether the wildlife fences are either single fences or double fences based on the characteristics of the wildlife fences. The evaluations were done during training and considered confusion matrix [62].

where t p stands for true positive, f p for false positive, and f n for false negative. Further, we compute the Mean Intersection-Over-Union (MIoU), a commonly used evaluation metric in determining the segmentation performance of an image [63]. The metric was computed as follows:

where c is the total number of classes, and IoU always ranges between 0 and 1. The ideal performance is 1, which means that the prediction is the same as ground truth values and vice versa for 0. When the intersection area is larger, IoU will be close to 1.

where ∣ g ∣ and ∣ p ∣ are the number of elements in g and p , respectively, and the symbol ∩ represents the intersection of two masks:

where r and p are recall and precision, respectively.

Macro-averaging, computes the average metric across classes without considering class frequency.

where n is the total number of classes, Metric i is the metric value of class i , and Metric is the metric (e.g., precision, recall, or F1-score) for class i .

The weighted averaging accounts for class frequency by adjusting the influence of each class on the overall metric

where n is the total number of classes, w i is the weight assigned to class i , Metric i is the metric value of class i , and M e t r i c is the metric (e.g., precision, recall, or F1-score) for class i .

7.1 Recognition of wildlife fences using aerial and still datasets

Table 4 shows the number of parameters for the different classification architectures used in this article. We observed that models with less parameters outperformed models with a bigger number of parameters. We also argue that models with smaller hard disk space are easily deployed on mobile clients compared to bigger models, which are usually slow, the CNN model achieved a classification test accuracy of 73% in aerial images and 82% in still images (Table 5). However, the ResNet model showed a better performance on the test set of 100% on aerial images and 95% on still images. This is justified by the transferable capabilities of ResNet from one adaptation domain to another, and their robustness to operate on small datasets. Results indicate that the CNN model performed poorly compared to ResNet. However, results also show that CNN performed better on aerial images than on still images, and ResNet performed competitively on both aerial and still images. Therefore, we recommend using transfer learning on the N2 wildlife fences dataset compared to CNN’s because transfer learning architecture performs better than CNN. In addition, the test accuracy using CNN was generally higher than train and validation accuracy. We attribute this to the small dataset used in this study and the inability of the CNN model to be robust compared to ResNet. Due to these weaknesses, we computed more metrics such as confusion matrix, precision, recall, F1-score, and support, with their associated average accuracy, macro average, and weighted average.

The confusion matrix in Figure 8(a)–(d) corresponds to the results from aerial images with CNN, aerial images with ResNet, still images with CNN, and still images with ResNet, respectively. Figure 8(a) shows 87.50 % double fences classified as double fences and 12.50 % double fence misclassified as single fences, as well as 57.14 % single fences classified as single fences and 42.86 % single fences classified as double fences. For Figure 8(b), we observe classifications of 100.00 % double fences and 100.00 % single fences with no misclassifications. Results of the confusion matrix when using still images with CNN and ResNet achieved acceptable results with 95.83 % double fences and 69.57 % single images predicted correctly as double fences and single fences, respectively. In the same Figure 8(c), 30.43 % single fences were classified incorrectly as double fences, and 4.17 % double fence is incorrectly classified as a single fence. Finally, Figure 8(d) results were from using still images and the ResNet model. The model achieved correct predictions of 100.00 % double fences and 91.30 % single fences. The model also misclassified 8.70 % single fences as double fences, with no misclassifications for double fences.

Figure 8

Four confusion matrix showing (a) images from drone classified using CNN, (b) images from drone classified using ResNet, (c) images from standalone camera classified using CNN, and (d) images from standalone camera classified using ResNet. The images the model made predictions on are the test data. Source: Created by the authors. (a) Aerial, CNN, (b) Aerial, ResNet, (c) Still, CNN, (d) Still, ResNet.

As in Table 6, the CNN model achieved a classification average accuracy of 73% on aerial images and 83% on still images. However, the ResNet model performance is of 100% on aerial images as well as 96% on still images. Generally, the classification models performed better on aerial images than still images because the drone can capture wildlife fences from the front view, back view, and top view, while the standalone camera captures only the front view. Meaning drones capture more details of the fence and its associated features as compared to a standalone camera. Based on these results, we recommend the use of transfer learning on the wildlife fences dataset as compared to CNNs when using classification algorithms. We further attribute the poor performance to the smaller datasets used for both aerial and still.

7.2 Prediction and detection of wildlife fence insulators

On average, 12 days and 2 h were spent on training the aerial and still datasets. During training of the model, the loss, MIoU, accuracy, and dice coefficient were computed on the training and validation datasets for each training epoch. The parameters that were in the final epoch while training are considered the training results. All the parameters of the training step were stored in a dictionary with keys and values, and those values are retrieved and used for plotting the graphs. These graphs were used to evaluate the performance of the model. The model was considered good if the epoch vs accuracy graph was exponential and if the epoch vs loss graph was a hyperbolic curve. Figures 9 and 10 showed the loss, MIoU, accuracy, and dice coefficient curves of the aerial and still datasets, respectively. Figure 11 shows the impact of batch size on model performance. In both datasets (Table 7), the loss curves show some slight oscillations when computing the loss, accuracy, and dice coefficient metrics; these were related to the set small mini-batch size. We note that the network fails to find the optimal convergence direction at the early stage of training. On observing the loss curves, gradually, the network stabilizes whenever we have a continuous network training, which obtains results in Table 7. We also note that the dice loss, often set as the negative Dice coefficient, measures the similarity between the prediction and truth. Reducing the negative Dice coefficient (Dice loss) corresponds to maximizing the Dice coefficient, meaning that the model’s predictions match the ground truth better. This aspect helps optimize for high overlap between predictions and actual target regions.

Figure 9

Loss, MIoU, accuracy, and dice curves of the aerial image dataset model. Source: Created by the authors.

Figure 10

Loss, MIoU, accuracy, and dice curves of the still image dataset model. Source: Created by the authors.

Figure 11

The impact of 16 batch size and 8 batch size on the performance of the U-Net model. Accuracy looks stable among the aerial and still datasets and along the 16 and 8 batch sizes. However, we see that the MeanIoU and Dice coefficients are penalized when using the Still dataset compared to the Aerial dataset. This is due to how strict the MeanIoU and Dice Coefficient metrics are, which are more precise measures of segmentation performance by focusing on overlap in the regions of interest. Source: Created by the authors. (a) Aerial, Params and (b) Aerial, Params.

7.3 Mask prediction accuracy

We applied the U-Net model to the wildlife aerial image dataset and wildlife still dataset, which are labeled as insulators and not insulators. Every image of the data is RGB and has 512 × 512 pixels resolution. The best MIoU were recorded for training, validation, and testing for both aerial and still datasets in Table 7. Adam with learning was adopted as an optimization algorithm [64].

Model prediction output is an image with each pixel in it related to a probability distribution that detects an area of interest. The output image size corresponds to the input image size. Binary pixel values were obtained by choosing a threshold of 0.5. Any pixel value below a specified threshold was set to zero, and value above the threshold was set to 1. We then multiplied the results of every pixel in an output image by 255 and got a black-and-white predicted mask. Further, for each predicted ground truth mask, we calculated the residuals. The smaller the residuals, the better the model performance, and the more the residuals, the poor the model performance, as shown in Figure 12.

Figure 12

Representation of actual images, ground truth masks, predicted masks, and residual masks on aerial dataset. Source: Created by the authors.

7.4 Insulator detection with OpenCV

OpenCV is an open-source library widely used for image analysis, recognition, and deep learning. The software is used exhaustively to recognize objects in photos and videos. Object detection is a subset of signal processing, big data, and machine learning focusing on identifying features in pictures and videos. In this study, images were labeled as positive and negative to train a cascade function, which was used to detect objects in images [65]. We detected insulators using OpenCV as shown in Figure 13 from the aerial dataset by finding contours defined by joining lines of all points at the image boundary with the same intensity. OpenCV uses a function findContours() to extract contours from an image that meets a specified threshold. In addition, a boundingRect() function was adopted to approximate rectangles of the blobs on an image to show the area of interest after getting the image’s contours or outer shape. This function returned four values, the x-coordinate, y-coordinate, width, and height of the blob. These values were used to draw a rectangle on the image part using pixel coordinates. A few pixels were added to these values to draw a slightly bigger box. These coordinates from mask images were inferred into the actual images to draw the borders around them. The cut-outs images and their corresponding mask images were passed through our OpenCV code to detect all the sources in the image. Later, these images were concatenated back to an original image of which borders were sketched around it, as shown in Figures 13 and 14, applied to the aerial and still images, respectively.

Figure 13

Detected insulators on wildlife electric fence at Amakhala private game reserve from the aerial dataset [50].

Figure 14

Detected insulators on wildlife electric fence at Amakhala private game reserve from the still dataset [50].

7.5 Discussion

WVC incidences are high in South Africa particulary near areas with wildlife and wildlife fences [66]. This prompts researchers to focus on measures to minimize and prevent wildlife collisions and ensure the safety of drivers, pedestrians, and other road users. It is therefore imperative that ecologists and wildlife conservationists in Africa work together to develop ways to detect WVC areas with high accuracy and low cost while minimizing the weaknesses of roadkill prediction using time series analysis, an approach currently adopted in road ecology in detecting roadkill areas. To assist the ecologists, this study automatically investigates the possibility of the U-Net segmentation algorithm, CNN, and ResNet in detecting and recognizing WVC areas. It does this by understanding the performance of CNN, ResNet, and U-Net trained on still and aerial wildlife fence and wildlife fence insulator datasets. Ultimately, for segmentation, it verifies if this was a good approach for detecting WVC fences by further drawing bounding boxes on the predicted insulators as shown in Figure 13. In addition, the ResNet and CNN recognize WVC areas using wildlife single and double fences shown on the confusion matrix in Figure 8 as well as precision, recall, F1-score, and support in Table 6. Generally, the accuracy of the models is correlated to the size of the dataset; the bigger the dataset, the higher the accuracy and vice versa [67,68]. This is un-realist in an ecological sense since large amounts of training data can not be easily produced by ecological researchers. For that purpose, we designed the U-Net model for smaller ecological wildlife fences insulator datasets and adopted ResNet for the classification task, which performs better than CNN.

Our work can uniquely combine ecological wildlife fences, wildlife fence insulator data, and deep learning techniques. The field of image recognition, especially when applied to wildlife fencing in detecting WVC areas, is still young. While most researchers have used similar methodologies, they have used them in different fields but not in detecting WVC areas using wildlife fences to minimize WVC.

Our results show that ResNet, CNN, and U-Net are successful within the ecological domain, although they also show some limitations. The algorithms, if deployed, might be a powerful and practical tool that helps ecologists extract ecological information from the wildlife fencing algorithm deployed.

7.6 Future scope

We offer a few possibilities for future work by combining traditional wildlife fencing, classification, and segmentation algorithms.

First, we acknowledge that our datasets were not normally distributed. They comprised only two privately owned game reserves located in the; Eastern Cape, South Africa. As a result, future work should repeat the data collection process, and this should include different provinces of South Africa and test the model’s generalization between the different identified locations. This will improve machine learning systems in generalizing their training data, which makes the model better in generalizing [69].

Second, deploying machine learning models to the client devices is vital, as periodically retraining the model from model staleness. Deployment has been successful in public policy [70], health [71], education [72], and banking [73]

Third, to further improve the classification accuracy of remote sensing, adopting image localization and detection algorithms might help to detect the exact location of an object in an image using algorithms such as histogram of oriented gradients, YOLO, Region-based convolutional neural networks (R-CNN), Fast R-CNN, and Faster R-CNN.

Fourth, building more images of the dataset using synthetic images will improve the performance of the model by exponentially increasing examples given to the model during training. This approach is widely used by U-Net with generative adversarial networks and is successful [74,75].

Fifth, models developed using aerial and still datasets may be continuously trained and deployed to minimize model staleness by using proactive training [76]. This approach will improve model performance and reliability by retraining the images the model provided as incorrect predictions.

Finally, we believe that the future of deep learning models relies upon developing an end-to-end wildlife fence detection system to monitor and detect potential WVC areas. The system should detect wildlife fences as WVC areas and alert the driver. Further, the system should allow the driver to save images and GPS. The collected image dataset can be used to retrain the model and the GPS will help improve the existing roadkill areas detection system by detecting future patterns and distribution of WVC areas using time series analysis.