Insect-pests can affect plants by disrupting or interfering with one or more physiological functions that lead to below-normal performance, such as reduced biomass and grain yield. Insect- pests can damage plants in several different ways, by killing plants, which can leave a gap in the crop stand, and the inability of plants to compensate for the open stand (e.g., some boring insects), general stunting caused by metabolic disruption through the nutrient drain or root damage (e.g., aphids, grubs). Moreover, they can kill branches (some species of scale insect can result in branch die-back) or eat inflorescence (e.g., beetles) or plant organs (e.g., boring insects), at-harvest or post-harvest losses (e.g., borers, weevils, etc) (Singh et al., 2021b; Higley, 1986; Pedigo et al., 2021). Insect-pests also cause damage by spreading diseases (Singh et al., 2021a).

Early detection, counting, and constant monitoring of the insects are vital to manage insect pressure in agriculture and reduce the pests’ infestation (Singh and Singh, 2005; Higley, 1986; Stern et al., 1959), as it helps farmers and agricultural professionals monitor and assess the population dynamics of various insect species within their fields. This information is important for making informed decisions about pest control strategies (Sarkar et al., 2023). By tracking the abundance of insects, farmers can identify potential outbreaks early on and take measures to prevent or mitigate crop damage. Moreover, establishing threshold levels helps determine when the insect population reaches a point where action (e.g., pest control) is necessary (Lima et al., 2020). For instance, if insect populations are increasing rapidly or reaching the action threshold, farmers can implement targeted pest control measures, such as applying insecticides or deploying predators, to prevent significant crop losses (Pedigo et al., 2021).

Furthermore, scouting for pests provides valuable data for integrated pest management (IPM) programs. IPM is a sustainable approach that aims to minimize the environmental impact of pest control while maximizing crop yields. Accurate insect counts help IPM practitioners determine the appropriate timing and intensity of pest control interventions reducing the reliance on broad-spectrum insecticides that can harm beneficial insects and lead to insecticide resistance (Cardim Ferreira Lima et al., 2020). Therefore, insect counting is essential in agriculture as it enables farmers to make data-driven decisions, minimize crop damage, and adopt environmentally friendly pest management practices, ultimately contributing to more sustainable and productive farming systems. Manual methods, such as analyzing sticky traps in the field to observe and quantify insects, are time-consuming and labor-intensive tasks and also requires human expertise in accurate pest identification. Therefore, a more automated insect detection and quantification method will be useful for plant researchers and farmers.

Earlier, insect detection was based on their differences in shape, color, pixel intensities, grayscale intensity, and texture analysis (Bauch and Rath, 2005; Huddar et al., 2012; Ghods and Shojaeddini, 2016). More efficient methods are needed to enable accurate and timely monitoring of large crop production areas, which currently demand significant time and labor. In this regard, AI systems, enabled by machine learning (ML) hold great promise for varied phenotyping, for example in disease phenotyping (Rairdin et al., 2022), yield estimation (Riera et al., 2021) and root traits (Jubery et al., 2021; Falk et al., 2020). Similarly, AI/ML is a necessary tool for automatic insect recognition from images and it has led to the development of several automatic monitoring systems (Barbedo, 2020; Li and Yang, 2020; Li et al., 2021; Rustia et al., 2021a; Wang et al., 2020d; Singh et al., 2021a). For instance, a large deep learning model was developed using citizen science data to detect a wide variety of insects (Chiranjeevi et al., 2023) ‘in the wild’ with high robustness Saadati et al. (2023). Apart from such large-scale models, researchers achieved a mean Average Precision (mAP) score of 63.54% using YOLOv3 Redmon and Farhadi (2018) to detect and classify pests in their “Pest24” dataset, which contained 25,378 annotated images of 24 pest species collected using an automatic imaging trap (Wang et al., 2020a, b, d). A multi-stage deep learning method that included object detection, insect vs. non-insect separation, and multi-class insect classification was proposed, achieving an impressive average F1-scores of up to 0.92 (Rustia et al., 2021b). Moreover, an AI-based pest counting method for monitoring the black pine bast scale (M. thungergianae) was developed, which reached a counting accuracy of 95%.

However, there are still challenges to address including data collection conditions (Hong et al., 2021) and hence, there is a lack of robust and field-ready insect monitoring systems (SmartProtect, 2022). Many existing studies use datasets with close-up, high-quality images that do not accurately represent the challenging field environments (Cheng et al., 2017; Kasinathan et al., 2021; Nanni et al., 2020; Pattnaik et al., 2020; Wang et al., 2017, 2020b). Building imaging systems automatically capturing high-quality snapshots of individual insects is difficult, especially for small or flying insects (Barbedo, 2020). Therefore, in many instances, a more practical approach is to capture a surface covered with multiple trapped insects using a single image within a sticky trap in the field (Ding and Taylor, 2016; Jiao et al., 2020; Rustia et al., 2021a; Xia et al., 2015; Zhong et al., 2018). Smaller tiles of individual insect images can then be extracted from the full image for further analysis.

Insect monitoring systems often focus on detecting a single pest, overlooking the potential presence of other species that could provide valuable ecosystem information (Ding and Taylor, 2016; Hong et al., 2021; Nazri et al., 2018; Roosjen et al., 2020). A recent study has shown that vision-language foundation models can be leveraged for zero-Shot (without requiring additional model fine-tuning) insect detection (Feuer et al., 2023). Additionally, many AI practitioners fail to apply strict validation procedures, leading to known methodological pitfalls like “data leakage” (Kapoor and Narayanan, 2022). Previous research has demonstrated that model performance can be overestimated when weak validation procedures, such as random data splitting, are used (Kalfas et al., 2021, 2022). Researchers applied different object detectors to localize and classify the insects simultaneously, such as YOLO, R-CNN, and Faster R-CNN (Li et al., 2021; Nieuwenhuizen et al., 2019). They also take advantage of transfer learning and initialize their model using the COCO dataset (Lin et al., 2014). Leveraging Transfer learning, it has been shown that YOLOv4 and YOLOv5 have relatively good performance in detecting five insect species (Verma et al., 2021). Self-supervised learning can also be effective for developing insect detection models (Kar et al., 2023). However, detecting insects on sticky traps using Deep learning (DL) still has some challenges, such as the lack of training data, the small size of insects, the similarity between different insect species, and significant morphological differences among stages in the life cycle of each species of insect, which makes the detection a complex task (Akintayo et al., 2018). Furthermore, some studies focus on datasets with broad insect classes, making classification relatively easier but not representative of more challenging scenarios (Rustia et al., 2021a; Wang et al., 2020d). To overcome these limitations, future research should address the practicality of capturing images in the field, consider the presence of diverse insect species, and implement rigorous validation procedures to ensure accurate and reliable insect monitoring systems.

In this paper, to address some of these challenges, especially the lack of data, we propose a machine learning framework to identify and localize pests on yellow sticky traps using a state-of-the-art object detector in the YOLO series called YOLOv8. On top of using YOLOV8, we leverage two techniques, namely Hierarchical Transfer Learning (HTL) and Slicing-Aided Hyper Inference (SAHI), to alleviate the issues due to smaller training data and small size of the objects of interest. HTL is an advanced version of traditional transfer learning, which leverages knowledge from a larger dataset to improve accuracy when training on a smaller dataset. It involves multiple steps of transfer learning, using intermediate datasets closely related to the target domain, known as in-domain datasets, to enhance the model’s learning process. This iterative approach allows the model to gain insights from datasets that share similarities with the target dataset, leading to significant improvements in performance. To further enhance the accuracy of detecting smaller-sized pests (e.g., an adult Western Corn Rootworm (WCR) Beetle is typically 1 4 inch long) in images, we implemented the Slicing Aided Hyper Inference (SAHI) (Akyon et al., 2022) method. It enhances the detection of tiny pests in images by dividing the original image into overlapping patches and independently subjecting each patch to object detection, improving overall performance. It also performs a full-inference step to detect larger objects, and then combines the results from both patch-wise and full inference using Non-Maximum Suppression (NMS) to ensure comprehensive and accurate object detection outputs. We report that using HTL instead of vanilla transfer learning, as in previous works, can improve detection accuracy significantly. The further addition of SAHI into our inference framework proves to be a useful strategy for detection of small insects-pests on yellow sticky traps. In addition, our choice of YOLOV8 lets us scale the YOLOv8 up and down to small and large networks and, at the same time, maintain the inference time and accuracy. Additionally, we report that using HTL instead of vanilla transfer learning, as in previous works, can improve detection accuracy significantly. HTL is an advanced version of traditional transfer learning, which leverages knowledge from a larger dataset to improve accuracy when training on a smaller dataset. It involves multiple steps of transfer learning, using intermediate datasets closely related to the target domain, known as in-domain datasets, to enhance the model’s learning process. This iterative approach allows the model to gain insights from datasets that share similarities with the target dataset, leading to significant improvements in performance. To further enhance the accuracy of detecting tiny pests in images, we implemented the Slicing Aided Hyper Inference (SAHI) Akyon et al. (2022) method to enhance the detection of tiny pests in images by dividing the original image into overlapping patches and independently subjecting each patch to object detection, improving overall performance. It also performs a full-inference step to detect larger objects, and then combines the results from both patch-wise and full inference using Non-Maximum Suppression (NMS) to ensure comprehensive and accurate object detection outputs. SAHI proves to be a valuable technique for object detection during inference.

Before using DL, data preprocessing and augmentation techniques were applied to increase the number of data samples and improve the model’s robustness. In our experimental setup, we employed a rigorous data splitting strategy to ensure the validity and robustness of our results. The dataset was divided into three distinct sets: training, validation, and test. Specifically, we allocated 80% of the data for training, 10% for validation, and 10% for the test set. This test set was kept completely separate and unseen by the model throughout the entire training and fine-tuning process. For training set, each image was tiled to 4 images; therefore, we could increase the data to 628 images. Moreover, before training, techniques such as rotating, zooming, flipping, changing illumination, and mosaic augmentation were applied to the images.

In our experimental framework, we orchestrated HTL trials, which comprised a two-step fine-tuning process. Initially, the COCO pre-trained model underwent fine-tuning on in-domain datasets, followed by a subsequent fine-tuning phase on our specific dataset. This design yielded three distinct experiment categories: (1) HTL: COCO-Kaggle, (2) HTL: COCO-IP102, and (3) COCO-Pest24, with the Kaggle (Yellow Sticky Traps), IP102, and Pest24 datasets respectively. The results, graphically illustrated in Figure 4 , unveil compelling insights through mAP and mAP50-95 plots. Notably, the COCO-IP102 experiment emerged as the most successful, excelling in both mAP50 and mAP50-95. This accomplishment can be attributed to the expansive diversity of insect species encapsulated within the IP102 dataset (contains >75,000 images belongs to 102 different insects), encompassing pivotal categories including beetles that align with our focus. While the Kaggle dataset’s limited size hindered it from surpassing Pest24 and IP102, its performance, as depicted in Figure 4B , showcased improvements over the baseline in terms of mAP50-95. We posit that a larger Kaggle dataset, given its close resemblance to our data, could potentially yield enhanced results. Despite Pest24’s abundant data, its divergent background and data characteristics pose challenges, thereby compromising insect details, particularly in comparison to the more distinct IP102 dataset. Furthermore, Figure 5 underscores precision and recall values. These results corroborate the important role played by IP102 as an in-domain dataset, wielding a marked influence in elevating precision and recall metrics. The discernible impact of in-domain datasets on precision is evidenced, reaffirming their role in augmenting overall performance.

The practical implications of our model’s performance are shown in Figure 6 , where the insect detection results are visually presented for two sample images. The result shows the model’s capability to detect small insects through the comprehensive synergy of HTL and the SAHI framework as shown in Table 1 . These outcomes were obtained from the HTL: COCO-IP102 experiment, which showcases the performance of the approach in the realm of HTL. Figure 7 illustrates the procedural enhancements facilitated by our method, which incorporates HTL and SAHI, aiming to enhance the detection of small insects.

Additionally, in Table 1 , the inference time for the model on the system with an Intel Core i7 processor and an NVIDIA T4 GPU averages around 0.06 seconds per image (640x640 resolution), with fluctuations ranging from 0.04 to 0.1 seconds. When incorporating the SAHI post-processing method, the inference time increases to an average of 0.7 seconds per image, ranging from 0.3 to 1.2 seconds depending on object density. On a Raspberry Pi 4B, equipped with a 1.8 GHz quad-core Cortex-A72 CPU and up to 8GB of RAM, the inference time, including post-processing, ranges from 8 to 15 seconds per image.

Recognizing the critical importance of insect pressure, we have conducted a comprehensive analysis to determine the insect populations in both predicted and ground truth scenarios. Our findings indicate that, when evaluated against the defined threshold for insect pressure, the predicted results closely align with the ground truth data. This alignment suggests that our predictive model can effectively identify the severity of insect presence, mirroring the accuracy of the ground truth measurements. This threshold serves as a valuable indicator for farmers, enabling them to make informed decisions about the application of insecticides when insect pressure surpasses the established threshold limit. It is worth noting sticky traps were put in the fields with the growing plants; we noted that other than insect-pests, plant parts were also stuck in the sticky traps (see Figure 7 ). Irrespective of this problem, the model could detect the majority of insects compared to the ground truth, thereby demonstrating its usefulness to plant scientists and farmer communities.

In this study, we extensively explored the efficacy of Hierarchical Transfer Learning (HTL) for the detection of insects on sticky traps under the constraint of limited training data. We demonstrate efficacy of HTL using three publicly available in-domain datasets. The HTL experiments underscore the importance of selecting effective in-domain datasets to optimize model performance. The two-step fine-tuning process revealed that the COCO-IP102 dataset, with its extensive diversity and volume significantly outperformed other datasets. This highlights the necessity of choosing datasets that not only align closely with the target application but also encompass a wide variety of classes and scenarios. Although the Kaggle-Yellow Sticky Traps dataset showed promise, its limited size restricted its performance relative to IP102. Conversely, while larger datasets like Pest24 offered abundant data, their less relevant characteristics hindered effective feature extraction. In conclusion, Prioritizing in-domain datasets that reflect the specific conditions and target species of the application, while also ensuring sufficient diversity and volume, is essential for enhancing performance in hierarchical transfer learning. This strategic approach can lead to more robust models capable of achieving superior performance in real-world applications.

To further improve the performance of detecting small sized insects, we consider the Slicing Aided Hyper Inference (SAHI) method; a strategic approach that capitalizes on image resolution that improved the insect detection capabilities. Finally, We present the design of a hardware setup and an efficient cyberinfrastructure for deploying the persistent insect monitoring framework in real life. While our study demonstrates the effectiveness of YOLOv8 combined with HTL and SAHI for insect detection on sticky traps, we acknowledge certain limitations in our approach. Due to the constraints of edge device deployment, which requires careful consideration of memory and computational resources, we were unable to explore more resource-intensive state-of-the-art methods such as transformer-based models. These advanced techniques, while potentially more powerful, are often impractical for deployment on resource-constrained devices. Future research could focus on adapting these models to enhance detection accuracy, should improve performance be a priority.

Sticky trap-based accurate early detection and counting allow for early mitigation of insect pests. This will allow stakeholders to precise control once they know which insect is trapped in a sticky trap and the level of insects based on the counting, allowing better management. The decision to spray insect pests before the pest population reaches economic injury level (EIL) will allow farmers to apply pesticides when the insect population has reached the action threshold. This will prevent using broad-spectrum insecticides indiscriminately, which helps to avoid pest resistance problems as well as leads to sustainable agriculture.