IoT and Machine Learning for Smart Bird Monitoring and Repellence: Techniques, Challenges, and Opportunities

1. Introduction

Birds cause damage to crops during the period before harvesting at a global level, leading to huge losses exceeding billions of dollars yearly [1]. The birds feed on seeds, fruits, leaves, and grains, reducing crop yields and threatening food security for many communities [2] . This not only impacts farmers’ livelihoods but also raises concerns about the availability of food for families relying on those crops [3]. In the East African region, small-scale farmers face a battle each season as bird infestations take a significant toll on their cereal crops, with losses exceeding 20 percent of the produce [4,5]. Farmers apply traditional bird control methods [6], which include the use of propane cannons, reflective tapes, and physical controls using nets and scarecrows [7]. Although these measures can lead to improved yields, their effectiveness is dependent on the stage of crop growth and duration of application [8]. The traditional methods are time-consuming, require intensive labour, and are mostly inefficient if not applied consistently throughout the day. In addition, methods that involve the use of chemicals result in environmental risks. The birds are also able to gradually adopt to static methods like the use of scarecrows, making them lose effectiveness over time. The practicability of using manual surveillance of large farms is also a problem. Given the challenges, the traditional bird control methods are inadequate for repelling birds in large-scale farms.

Recent advances in Artificial Intelligence (AI) [9], specifically in computer vision and machine learning, supported by the growing adoption of the Internet of Things (IoT) technologies [10,11] present opportunities for automating bird detection and repellence [12,13]. These technologies can be used to detect, classify, and respond to bird activity in real time using edge computing devices, such as drones, camera traps, and smart sensors [14,15,16]. In addition, AI-powered repellence techniques, which may include adaptive sound emitters, automated lasers to predator-mimicking drones, are gaining popularity as more effective alternatives to the traditional deterrent methods [17,18]. Despite growing interest and the need for a real-time solution, the research landscape on AI and IoT-based bird detection and repellence remains fragmented [19]. Previous studies focus on isolated components such as classification models, wireless sensor deployment, or acoustic deterrants without synthesising the full pipeline, from data collection and model selection to edge deployment and deterrence actuation [20,21,22]. The growing volume of literature presents a challenge for practitioners and researchers trying to navigate the field.

As an initial step to addressing the gaps in this area of study, this paper presents a systematic review of recent studies (2020–2024) in AI-enabled bird detection and deterrence. Using the PRISMA framework, 154 peer-reviewed studies that leverage machine learning, computer vision, and IoT infrastructure for avian monitoring were identified and analysed from the initial 267. Unlike prior reviews [7,23,24], this study evaluates the full pipeline—from data collection and model architecture to deployment platforms and smart deterrence technologies. This review not only classifies the models and architectures used but also evaluates their real-world deployment readiness, data requirements, edge processing strategies, and integration with repellence technologies.

The key contributions of the study are as follows:

• Machine learning techniques applied in bird detection are categorised and mapped, highlighting trends in lightweight models and edge compatibility.
• Dataset types, collection methods, and preprocessing techniques used in training detection models are reviewed.
• IoT architectures and communication protocols are evaluated, identifying strengths and limitations in cloud-based systems.
• An analysis of bird repellence methods and their integration with intelligent detection systems is provided.
• Key challenges are identified, and future research directions for building scalable, adaptive bird management systems are proposed

By addressing these gaps, this review aims to support researchers, developers, and policymakers in designing effective, AI-powered solutions for sustainable bird monitoring and control.

2. Materials and Methods

The study followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework to review the literature on IoT and machine learning for bird detection and repellence. This ensured a replicable and unbiased selection of relevant studies. The process consisted of four main steps: Identification, Screening, Eligibility, and Inclusion.

2.1. Identification

The first step was identifying relevant studies. A comprehensive search was conducted in IEEE Xplore, ScienceDirect, Springer, and Google Scholar, covering studies published after 2020. The search included terms such as:

• "Machine Learning + Bird Detection"
• "Machine Learning + Bird Repellence"
• "Computer Vision + Bird Detection"
• “Acoustic Bird Detection”
• "IoT + Bird Repellence"
• "Artificial Intelligence + Bird Detection + Repellence"

These search terms helped capture a wide range of studies, from advanced edge computing applications to real-world agricultural use cases. This initial search retrieved 267 articles for potential inclusion.

2.2. Screening

With the initial pool of studies collected, the next phase was the screening phase. First, automated tools were used to eliminate duplicate entries, reducing the dataset to 253 papers. Then, a manual screen of the titles and abstracts was done to ensure that only studies directly related to our research topic were considered. Studies were excluded if they lacked machine learning applications, did not involve IoT technologies, or focused on general wildlife monitoring without specific reference to birds. After screening, 248 papers were identified for potential review, out of which 10 were not retrieved.

2.3. Eligibility

The full texts of the remaining 238 papers were manually reviewed to check if they met our predefined inclusion and exclusion criteria as presented in Table 1.

2.4. Inclusion

After applying the eligibility criteria, 154 papers were selected for data extraction and analysis. These studies provided insights into various aspects of IoT-based bird detection and repellence, including dataset characteristics, machine learning techniques, hardware implementations, model performance, and real-world applications. The extracted data focused on key themes such as Datasets,machine learning algorithms, IoT architectures, Connectivity, edge-based deployments, bird repellence techniques, and implementation challenges. The findings were synthesized to identify trends, gaps, and opportunities for further research.

Figure 1. Study selection steps.

Figure 1. Study selection steps.

3. Computer Vision-Based Detection

3.1. Datasets

The reliability and accuracy of AI models for bird detection and monitoring largely depend on the quality and diversity of the datasets used [25]. Bird datasets are highly dynamic, as birds move across varying environments with changes in lighting, weather, and habitat conditions. Therefore, researchers need large, well-annotated datasets that represent different species, behaviours, and environmental factors to build effective detection systems [26,27]. The datasets used are often visual or sensor-based, with some models employing a combination of two or three different datasets to enhance model performance [28]. Collecting the necessary data requires careful selection of collection methods, data sources, and preprocessing techniques to ensure the datasets are not only comprehensive but also useful. Table 2 gives a summary of the analysed datasets.

Different data collection methods have been applied in various bird detection studies depending on the research goals and the available technology. The most common methods from the reviewed studies include: video surveillance [31], Image-based methods [29], motion detection [37] and Multi-Sensor approaches [54]. The data used is collected mainly by use of cameras, which are considered to be affordable and easy to deploy, PIR Sensors for automated detection systems, drones, radar, ultrasonic sensors, and microphones. Publicly available datasets and GPS data are also used for training models and to allow researchers to cross-validate findings with existing data.

The size of a dataset can impact the model’s performance. In the reviewed studies, dataset sizes ranged from fewer than 1,000 images to over 300,000 samples. Some studies reported dataset sizes in alternative formats, such as hours of video footage rather than individual image counts. Most studies relied on custom datasets [50], demonstrating the need for specialised data collection. Only a few studies use widely available datasets such as the Kaggle datasets and the COCO dataset. Other specialised datasets were also used, for example, NIPS4Bplus, Xeno Canto, and warblrb10k. The heavy reliance on custom datasets suggests that existing datasets may not always meet the specific requirements of bird detection models, especially when dealing with regional species or unique environmental conditions. While large datasets improve model robustness, many studies still rely on relatively small collections, making data augmentation essential. The lack of publicly available datasets suggests a strong need for more open-source contributions to the field.

Raw data alone is rarely sufficient for training machine learning models. Researchers apply preprocessing techniques to improve data quality and enhance model accuracy. The most commonly reported preprocessing methods were: Annotation [34], data augmentation [38], image resizing and scaling [45], frame extraction [39], and feature extraction [53] The variety of data collection methods, dataset sizes, and preprocessing techniques indicates that bird detection research is still evolving.

3.2. Machine Learning Models

Machine learning models come in different architectures, each designed for specific strengths in detecting, classifying, and tracking objects. The studies reviewed used a variety of models, with some being applied in most studies. Table 3 provides a summary of the ML models used in different studies.

CNNs were the most frequently used model type. CNNs work by detecting patterns such as edges, textures, and shapes layer by layer, making them highly effective for image recognition tasks [66,67,68]. Variants like ResNet (Residual Networks) enhance CNNs by allowing deep networks to train more effectively without losing important details [69]. YOLO (You Only Look Once) models appeared in many studies [70,71], too. Unlike CNNs, which process an image in sections, YOLO treats the entire image as a single input, enabling real-time object detection [72] . Several versions of YOLO were used in the reviewed studies [73], with improvements in detection speed and accuracy. Faster R-CNN is widely recognised for its high accuracy in object detection tasks. Unlike YOLO, which prioritises speed, Faster R-CNN processes images in multiple steps, refining its predictions to improve precision [74]. This makes it a better choice for tasks where detection accuracy is more important than speed. MobileNet is used in low-power, edge-based applications. Unlike traditional CNNs, which require significant computational power, MobileNet is optimised to run on mobile devices, IoT sensors, and embedded systems [75]. VGG, Inception, and EfficientNet have deep learning capabilities but tend to require high computational resources and are not frequently used in bird detection. Traditional models like K-Nearest Neighbors (KNN), Hidden Markov Models (HMM), and Support Vector Machines (SVM) have also been explored; these methods are often used as benchmarks but are generally less effective for large-scale image analysis [76]. Some studies experimented with combinations of architectures, demonstrating that integrating multiple models can improve performance [77].

Accuracy was the most commonly reported metric, with the performance breakdown showing that most models achieved strong results, with most studies reporting accuracies ranging from 80-95 percent, confirming the effectiveness of deep learning models for detection and classification. In addition, precision was used to measure how many of the detected objects were correct, with studies reporting precision above 0.080-0.90. Recall was also used for measuring how well models detected all relevant objects in an image, with values ranging from 0.65 to above 0.95, with higher values indicating fewer missed detections. Mean Average Precision (mAP) was used, offering a balanced view of precision and recall. The reported mAP ranged between 70-9 percent, meaning highly effective detection. Frames Per Second (FPS) was reported in 6 studies. The reported speeds are between 1-60 FPS, which is still usable but may not be ideal for fast-moving objects. Figure 2 gives a comparison of the precisions from different models,

As AI moves towards real-world applications, the ability to run models on edge devices (such as Raspberry Pi, smartphones, or IoT sensors) is becoming increasingly important. Some studies reported edge-compatible models, showing that lightweight architectures like MobileNet and optimized CNNs are becoming more viable for real-world use.

4. Acoustic-Based Detection

Table 4. Acoustic-Based Detection

Study Focus	Hardware Used	Approach	Detection Performance
Evaluation of BirdNET for detecting two bird species [78]	AudioMoth	BirdNET (CNN-based)	Precision: 92.6% (Coal Tit), 87.8% (Short-toed Treecreeper)
Bird sound Classification [48]	No mention found	Multilayer Perceptron (MLP)	Accuracy: 74%
Vineyard protection from birds [79]	Raspberry Pi 3B, microphone	Two-phase: SVM and CNN	Accuracy: 96%
BirdCLEF 2021 challenge [80]	No mention found	CNN-based ensemble	F1 score: 0.6780
Birdsong detection on IoT devices [81]	STM32 Nucleo H743ZI2 MCU	ToucaNet and BarbNet (CNN-based)	AUC: 0.925 (ToucaNet), 0.853 (BarbNet)
Acoustic bird repellent system [82]	Arduino Nano 33 BLE, microphone	DenseNet201 (CNN)	Accuracy: 92.54%
Avian pest deterrence [83]	Arduino Nano 33 BLE Sense, XIAO ESP32S3	Conv1D neural network	Accuracy: 92.99%
Bird song recognition on IoT devices [84]	ARM Cortex-M microcontrollers	Various CNN and Transformer models	Accuracy: >90% for best models
Avian diversity monitoring [85]	Autonomous Recording Units (ARUs)	BirdNET (ResNet-based)	mAP: 0.791 for single-species recordings
Monitoring Eurasian bittern [78]	AudioMoth	BirdNET and Kaleidoscope Pro	Accuracy: 93.7% (BirdNET), 98.4% (Kaleidoscope Pro)
Passive acoustic monitoring of bird communities [86]	SM4 Wildlife Acoustics ARUs	CNN (ResNet50)	mAP: 0.97
Detecting novel bird species and individuals [87]	No mention found	Variational Autoencoder (VAE)	FPI: 1.6%, FNI: 0.9% (species detection)
Birdcall identification on embedded devices [77]	Jetson Nano	CNN-based multi-model network	Accuracy: 84.9%
Endangered birds monitoring [88]	ARM Cortex M3 micro-controller	Dynamic Time Warping (DTW)	No mention found
Bird species monitoring and song classification [49]	5G IoT-based system, ESP32-S3 MCUs	Various CNNs (EfficientNet, MobileNet)	Accuracy: >70% for best models
Evaluation of acoustic recorders and BirdNET [89]	AudioMoth, Swift Recorder, SM3BAT, SM Mini	BirdNET (not specified)	Accuracy: 96%
Bird audio detection [90]	No mention found	Lightweight CNN	Accuracy: 86.42%
Acoustic monitoring of avian species [91]	AviEar (IoT-based wireless sensor node)	No clear mention found	Precision: 99.6%, Recall: 95%

Table 4. Acoustic-Based Detection

Study Focus	Hardware Used	Approach	Detection Performance
Evaluation of BirdNET for detecting two bird species [78]	AudioMoth	BirdNET (CNN-based)	Precision: 92.6% (Coal Tit), 87.8% (Short-toed Treecreeper)
Bird sound Classification [48]	No mention found	Multilayer Perceptron (MLP)	Accuracy: 74%
Vineyard protection from birds [79]	Raspberry Pi 3B, microphone	Two-phase: SVM and CNN	Accuracy: 96%
BirdCLEF 2021 challenge [80]	No mention found	CNN-based ensemble	F1 score: 0.6780
Birdsong detection on IoT devices [81]	STM32 Nucleo H743ZI2 MCU	ToucaNet and BarbNet (CNN-based)	AUC: 0.925 (ToucaNet), 0.853 (BarbNet)
Acoustic bird repellent system [82]	Arduino Nano 33 BLE, microphone	DenseNet201 (CNN)	Accuracy: 92.54%
Avian pest deterrence [83]	Arduino Nano 33 BLE Sense, XIAO ESP32S3	Conv1D neural network	Accuracy: 92.99%
Bird song recognition on IoT devices [84]	ARM Cortex-M microcontrollers	Various CNN and Transformer models	Accuracy: >90% for best models
Avian diversity monitoring [85]	Autonomous Recording Units (ARUs)	BirdNET (ResNet-based)	mAP: 0.791 for single-species recordings
Monitoring Eurasian bittern [78]	AudioMoth	BirdNET and Kaleidoscope Pro	Accuracy: 93.7% (BirdNET), 98.4% (Kaleidoscope Pro)
Passive acoustic monitoring of bird communities [86]	SM4 Wildlife Acoustics ARUs	CNN (ResNet50)	mAP: 0.97
Detecting novel bird species and individuals [87]	No mention found	Variational Autoencoder (VAE)	FPI: 1.6%, FNI: 0.9% (species detection)
Birdcall identification on embedded devices [77]	Jetson Nano	CNN-based multi-model network	Accuracy: 84.9%
Endangered birds monitoring [88]	ARM Cortex M3 micro-controller	Dynamic Time Warping (DTW)	No mention found
Bird species monitoring and song classification [49]	5G IoT-based system, ESP32-S3 MCUs	Various CNNs (EfficientNet, MobileNet)	Accuracy: >70% for best models
Evaluation of acoustic recorders and BirdNET [89]	AudioMoth, Swift Recorder, SM3BAT, SM Mini	BirdNET (not specified)	Accuracy: 96%
Bird audio detection [90]	No mention found	Lightweight CNN	Accuracy: 86.42%
Acoustic monitoring of avian species [91]	AviEar (IoT-based wireless sensor node)	No clear mention found	Precision: 99.6%, Recall: 95%

As presented in table 4, the most common hardware used for audio data collection is the AudioMoth and ARUS, and other sound sensors. The Audio moth has a detection range of 801-900m, while other devices have a range of below 200m, making an Audio moth appropriate for large-scale projects. The microcontrollers and processors used include: STM32, Arduino, ESP32, and ARM, with Raspberry Pi and Jetson Nano also being used in a few studies.

The CNN-based models including variants like ResNet were commonly used with other approached including; MLP, SVM, Transformer, VAE, and DTW

5. Connectivity

Once data is captured, it must be transmitted efficiently. Studies reviewed indicate a mix of wired and wireless communication protocols, with a strong preference for wireless due to flexibility and scalability. The common wireless communication technologies used include;

• Wi-Fi - This enables high-speed data transfer and has been applied in several studies. However, it has a limited range and high power consumption, making it unsuitable for large-scale, battery-powered networks.
• LoRa (Long Range, Low Power) – This has also been used and is ideal for IoT applications in agriculture and environmental monitoring due to its long range and low power needs. However, the low data rate make it less suitable for applications requiring high-resolution image or video transmission.
• Cellular Networks (4G/LTE, 5G) – This has been used to provides seamless connectivity, especially for mobile IoT devices. However, high cost and energy consumption make it impractical for many large-scale IoT applications.
• Zigbee - Very low power consumption, low cost, well-suited for mesh networks in local IoT setups. Shorter range compared to LoRa and Cellular, not suitable for high-data applications like images or videos

No single communication technology meets all IoT requirements. Studies highlight trade-offs between long-range connectivity, power efficiency, and data transfer speed. Hybrid communication approaches for example combining LoRa for low-power sensing and Wi-Fi for bulk data uploads can optimize performance. Table 5 presents a comparison of the connectivity options

6. IoT Implementation Architectures

The way sensor data is processed and stored significantly impacts system efficiency. The reviewed studies revealed two dominant architectures:

6.1. Cloud-Based Architectures

Sensors send raw data to cloud servers for storage, processing, analysis for long-term decision-making [92]. This architecture is scalable and supports advanced machine learning models but requires high bandwidth requirements and has high latencies in case of poor connection.

6.2. Edge Computing

Sensors transmit data to a nearby edge device (e.g., Raspberry Pi, NVIDIA Jetson) for local processing before sending key insights to the cloud. The use of this architecture reduces latency and bandwidth usage and is ideal for real-time applications [93]. The edge devices have limited processing power and storage capacity. Bird detection systems have increasingly leveraged edge computing to enable real-time, efficient, and autonomous monitoring [94,95,96]. By processing data closer to the source—on the edge—these systems can reduce latency, minimize bandwidth usage, and operate effectively even in remote environments. Various edge devices have been explored in bird detection studies;

• Microcontrollers (ESP32, ATmega328, etc.) – These low-power devices are ideal for lightweight processing tasks but struggle with deep-learning models due to limited computational capacity [97].
• Single-board computers (Raspberry Pi, Jetson boards) – More powerful than microcontrollers and are commonly used in edge-based implementations, these devices can handle more complex computations but consume more power and are more costly [30].
• FPGA-based solutions – While highly efficient for real-time processing, FPGA implementations are less common due to their complexity and cost. Deploying machine learning models at the edge requires balancing of performance, power efficiency, and resource constraints. The reviewed studies explored several optimisation strategies:
• Lightweight models – MobileNet and optimized YOLO variants are frequently used due to their efficiency in object detection tasks.
• Transfer learning – Adapting pre-trained models allows for reduced computational overhead while maintaining high accuracy [98].
• Model compression – Techniques such as pruning and quantization help shrink models to fit within resource-limited devices [99].

TinyML—machine learning optimized for microcontrollers—has emerged as a promising approach for bird detection in energy-constrained environments. Studies have explored several techniques to make TinyML viable; Employing partial convolution and quantization to optimize TinyML models [84], and a lightweight CNN with fewer than 100,000 parameters, reducing memory consumption [90]. To maximize efficiency, TinyML-based bird detection relies on:

• Pruning and quantization – Reducing model complexity without significantly impacting accuracy.
• Power-saving techniques – Using sleep modes and efficient RAM allocation in microcontrollers.
• Local data processing – Minimizing the need for network communication to save power.

7. Bird Repellence Methods

Bird deterrence is an essential aspect of bird detection, especially in agricultural, conservation, and aviation settings [100]. Various techniques exist to prevent birds from interfering with crops, equipment, or infrastructure. These methods range from simple sound-based solutions to advanced AI-driven adaptive deterrence as presented in Table 6;

Sound-Based deterrents use loud noises, ultrasonic frequencies, or bioacoustic calls (e.g., distress signals of birds) to scare birds away. These works well initially, but birds may habituate over time, reducing the long-term impact . Visual deterrent methods include flashing lights, predator-mimicking drones, bird-scaring lines, and laser systems. These can be highly effective, especially when mimicking natural predators, but practical limitations exist for large areas. Integrated Approaches by combining sound, visual, and AI-driven adaptive systems enhance long-term effectiveness. These are promising, but studies suggest more long-term trials are needed.

8. Discussion

8.1. Challenges in Bird Detection and Repellence Systems

8.1.1. Detecting small and distant birds with high accuracy:

Birds, especially smaller species, are difficult to detect at long distances, making early intervention challenging. Factors like bird size, movement speed, and background complexity (e.g., sky, trees, or buildings) make accurate identification difficult. In critical environments such as airports and farms, where early detection is crucial, this limitation can lead to increased risks of bird strikes or crop damage.

8.1.2. Environmental Variability and Real-time Adaptation:

Bird detection models must work reliably under changing environmental conditions—rain, fog, night-time, and varying lighting conditions affect sensor performance. AI models often struggle with fluctuating backgrounds, leading to misclassification or lower accuracy in non-ideal conditions. Ensuring real-time adaptability while maintaining robustness in diverse weather conditions remains a major hurdle.

8.1.3. Energy Efficiency and Computational Constraints on Edge Devices:

Many bird detection systems rely on IoT devices in remote areas with limited power. Running deep learning models on low-power hardware like microcontrollers and single-board computers requires balancing model complexity, computational efficiency, and battery life. Power constraints also limit high-resolution image processing and continuous monitoring, making optimization essential.

8.1.4. Managing Data Collection, Storage, and Transmission:

Bird detection models require high-quality training datasets, but collecting and labeling diverse bird species across different regions is resource-intensive. Furthermore, high-resolution images and continuous video streams generate large amounts of data, creating challenges in real-time storage, bandwidth use, and cloud-based processing in remote areas. Efficient data compression and transfer strategies are needed to reduce costs while maintaining accuracy.

8.1.5. Reducing False Positives and Enhancing Species-Specific Identification:

Distinguishing birds from other airborne objects, such as drones, insects, or even moving tree branches, is challenging. High false positive rates can trigger unnecessary responses, while false negatives can lead to system failures. Additionally, different bird species may exhibit unique behaviors that influence detection accuracy. Models must be adaptable and capable of species-specific identification to ensure effective repellence measures.

8.2. Opportunities with AI in Bird Detection

8.2.1. Deploying Low-Power, AI-Driven Edge Computing Solutions:

Advances in TinyML allow models to run efficiently on low-power microcontrollers, enabling bird detection in remote areas with limited energy access. These edge-based systems reduce dependency on cloud computing, improving real-time detection and decision-making while minimizing energy consumption.

8.2.2. Multi-Sensor Fusion for Enhanced Detection Accuracy:

Combining visual data from cameras with complementary sensors—such as acoustic analysis for bird calls, motion sensors, and infrared imaging—improves identification accuracy. This multi-modal approach helps overcome limitations presented by varying lighting conditions and environmental noise, leading to more reliable detections.

8.2.3. Adaptive AI Models for Self-Learning and Context Awareness:

On-device learning and fine-tuning models for local conditions allow systems to continuously adapt to different bird species and environmental changes. AI-driven adaptive repellence methods could adjust strategies based on bird behavior, preventing habituation and improving long-term effectiveness.

8.2.4. Energy-Efficient Model Optimization for Scalability:

Techniques like model quantization, pruning, and knowledge distillation enable running complex AI models on resource-constrained devices. This optimization not only enhances real-time processing but also makes large-scale deployment in agricultural and urban settings more cost-effective.

8.3. Future Research Directions

To further advance AI-driven bird detection and repellence systems, future research should focus on:

• Developing ultra-lightweight, high-accuracy AI models – Improving TinyML capabilities to maintain performance while reducing computational demands.
• Enhancing automated data collection and labelling – Creating standardized, open-source datasets for training and benchmarking bird detection models.
• Designing self-learning AI models – Implementing on-device adaptation to reduce reliance on cloud retraining and improve real-time responsiveness.
• Exploring AI-driven, species-specific repellence techniques – Using behavior-based deterrence strategies that dynamically adapt to different bird species.
• Integrating bird detection into broader smart agriculture and urban management systems – Ensuring AI-driven bird monitoring complements existing environmental and precision farming technologies.

9. Conclusions

This review highlightes the growing potential of integrating machine learning and IoT technologies for smart bird detection and repellence. The studies reviewed illustrate progress in applying advanced computer vision and acoustic models for accurate bird identification across diverse environments. The adoption of edge computing and TinyML frameworks further demonstrates the feasibility of deploying real-time, energy-efficient solutions in remote and resource-constrained areas. Multi-modal sensor fusion and adaptive AI-driven repellence strategies offer promising directions for increasing system robustness and effectiveness. Despite these advancements, key challenges remain. These include limited availability of standardized datasets, species-specific detection issues, environmental variability, power constraints, and the need for scalable, low-latency deployment architectures. Addressing these issues will require interdisciplinary collaboration, innovation in low-power AI model design, and the development of open-access datasets tailored to ecological and agricultural contexts. Future research must focus on building intelligent, self-adaptive systems that can evolve with changing environmental conditions and bird behaviors. Integrating these solutions within broader smart agriculture and urban management ecosystems will be critical for sustainable environmental stewardship, improved crop protection, and minimized human-wildlife conflict.