Image Augmentation Using Both Background Extraction and the SAHI Approach in the Context of Vision-Based Insect Localization and Counting

Abstract

Insects play essential roles in ecosystems, providing services such as pollination and pest regulation. However, global insect populations are in decline due to factors like habitat loss and climate change, raising concerns about ecosystem stability. Traditional insect monitoring methods are limited in scope, but advancements in AI and machine learning enable automated, non-invasive monitoring with camera traps. In this study, we leverage the new Diopsis dataset that contains images from field operations to explore an approach that emphasizes both background extraction from images and the SAHI approach. By creating augmented backgrounds from extracting insects from training images and using these backgrounds as canvases to artificially relocate insects, we can improve detection accuracy, reaching mAP50 72.7% with YOLO10nano, and reduce variability when counting insects on different backgrounds and image sizes, supporting efficient insect monitoring on low-power devices such as Raspberry Pi Zero W 2.

1. Introduction

Insects are not only the most diverse group of organisms on Earth but are also integral to ecological functions that support human well-being, including pollination, pest regulation, and nutrient cycling. In [1], the authors conducted a pivotal study revealing a dramatic decline of 75% in flying insect biomass over 27 years within protected areas, highlighting the pressing need for conservation efforts to combat biodiversity loss. This finding underscores the ecological significance of monitoring insect populations to better understand and mitigate these declines. As insect populations decline worldwide, largely due to habitat loss, climate change, and pesticide exposure, the resulting disruptions to these ecological services could have far-reaching impacts [2]. Monitoring these populations is essential to conservation efforts, yet traditional approaches—such as manual surveys—are labor-intensive, costly, and limited in scope. Advances in technology, especially in AI and machine learning, have opened new avenues for efficient insect monitoring through automated systems that incorporate sensors, camera traps, and image recognition software [3]. One way, especially as applied to ecology and biodiversity assessment where insects are not typically terminated but only observed, is through vision-based configurations. For instance, autonomous camera settings (a tableau, a camera, an attracting light, and wireless connectivity), combined with AI algorithms (transformers and deep learning mainly), can detect, classify, and track insects across multiple environments, providing a non-invasive way to gather population data on a large scale [4]. In [3,4], the authors support the idea that deep learning and computer vision have transformative potential for entomology, enabling rapid, automated species identification and population monitoring on unprecedented scales. Their findings emphasize the role of AI in addressing biodiversity challenges. Despite these innovations, challenges remain, as the vast diversity and morphological similarity among insect species complicate accurate classification.

The synergy of AI and vision-based traps as applied to the assessment of insect biodiversity as well as in the agricultural sector have received much attention. In [5], the author reviewed the application of camera-equipped automated traps in a related case, identifying the benefits and limitations of current monitoring systems. This work highlights the practical applications and future directions of automated monitoring in agriculture. It was made obvious that embedding small deep-nets in low-power devices has some distinct benefits over uploading images to a server. In [6], the authors developed and embedded a YOLOv8 and SAHI Model framework, demonstrating how the SAHI framework, which we also employ, collaborating with object detection can address insect recognition and counting. This system provides insights for applying TinyML in ecological and entomological monitoring. In [7], the authors conducted a comparative study on camera- and sensor-based traps, offering a comprehensive evaluation of current monitoring technologies in pest management. Their work informs the design of optimized systems for insect detection and monitoring. This work continued in [8], where they reviewed advancements in insect pest traps and deep learning-based pest detection, highlighting the development of smart traps capable of identifying pest species autonomously. Their review identifies key technologies and methodologies that support efficient, scalable pest management. Their efforts for low-power embedded systems continued in [9], where they discussed a low-power, edge-based monitoring system for insect detection showcasing the potential of lightweight, field-deployable devices for agricultural pest detection and ecological monitoring. In [10], the authors explored the convergence of embedded systems, edge computing, and machine learning in their study on the Internet of Intelligent Things, which facilitates real-time environmental monitoring and data processing in remote settings.

In [11], the authors explored edge computing for urban insect traps, demonstrating how edge AI can support smart city applications by enabling efficient, real-time insect monitoring in urban areas. In [12], the same team introduced an innovative insect counting method embedded in e-traps that learn without manual annotation. This system enhances efficiency by automating image-based insect recognition, offering a sustainable solution for real-time monitoring.

In line with our key idea that the background is quite important in image-based insect recognition and counting, in [13], it was demonstrated that the color of traps has a significant impact on the detection accuracy of deep learning models used for insect recognition on sticky traps. This research highlights the need to consider environmental and design factors in insect monitoring setups.

The Diopsis dataset, available through the Arise Challenges platform, serves as a vital resource for insect monitoring research, providing a comprehensive, annotated image dataset for species recognition and detection tasks [14]. This data facilitates advancements in vision-based insect detection systems. The classification part of the Diopsis set has already been used in literature. For classifying insects on a multi-level taxonomy, see [15]. In their study, they introduced TaxaNet, a deep learning framework utilizing a hierarchical loss function to improve insect classification accuracy across various taxonomic levels. This work exemplifies how structured loss functions can enhance model performance in complex ecological datasets. In our work, we deal with the other part of the database, namely the Detection corpus.

Our main contribution is the idea of not focusing on the insect but on its background (i.e., the canvas). We took pictures from the training data, and we removed the insects, therefore extracting the pallets of the database. During training, we relocated the insects of the training data onto new canvases as a means of data augmentation that reduces the variability of detection results due to the background. In addition, we utilized the newly developed YOLO10 nano model on images of various dimensions and integrated it into low-power edge devices for autonomous insect counting. This approach achieves an mAP50 of 72.7% with a processing time of approximately 76.37 s per image. Although this processing time is relatively long, it is not a critical factor, as field-deployed insect counting systems typically transmit data only once per day.

2. Methods

The Diopsis dataset is a set of images originating from vision-based traps operated in the field and their corresponding annotation files [14].

In the detection part of the competition, the organizers provided participants with a complete annotated dataset consisting of 3965 images (resolution 3280 × 2464 pixels) with a total of 48,216 box-labeled insects. The public part of the dataset comprises a total of 3237 annotated images with the coordinates of the insects, though a subset of images is partially annotated. This database part’s primary task is to count insects in the realm of real-world difficulties that include intense variability on the background canvas on which insects lay. This reflects some of the scenarios that one comes across when dealing with real-life operations and not curated insect collections commonly found in museums and photo albums. The manifestation of this variability includes pictures from malfunctioning devices, images that are irrelevant to the insect detection task, obstruction from flowers, intense variability in shadowing, monochromatic images, illumination variations, and glaze due to the luring light used by traps (see Figure 1). In field operations, one cannot anticipate all unexpected situations that may occur to a device operating on long-time deployments, but operational systems need to be able to deal with such cases.

2.1. Image Preprocessing

2.1.1. Train–Test Split

We created a python script that randomly isolated 10% of the images. This set of 323 images was used as a test control set and is a set of images that our algorithms had never seen during the training process. We removed 20 images that were very poorly labeled by the Diopsis Challenge so the final set of images for the results control set consisted of 303 images. For all counting approaches, we used 5-fold cross validation. Each fold is composed of 80% training data and 20% randomly selected, non-overlapping splits. The annotation set was first transformed to the YOLO annotation file structure. In [16], the authors introduced version-10 as an end-to-end object detection model that achieves real-time performance while maintaining high detection accuracy. This model is especially relevant for resource-constrained environments, providing efficient object recognition capabilities. In the following experiment, we make use of one of the smallest models, the so-called nano, and our intention is to embed the architecture in autonomous devices.

Using YOLOv10nano in embedded applications for insect recognition and counting can be justified based on several key factors, considering the constraints that are typical in agricultural monitoring scenarios: 1. Computational Efficiency: YOLOv10nano is designed to run on resource-constrained devices such as Raspberry Pi Zero 2, ESP32, and other microcontrollers with limited memory and processing power. Its lightweight architecture allows real-time or near real-time inference, enabling continuous insect monitoring without the need for cloud processing. 2. Low Power Consumption: Its minimal resource usage and power efficiency results in reduced power consumption, making it suitable for battery-operated devices in remote agricultural environments. The reduced power demand allows for longer operation times, less frequent battery replacements, or the use of small solar panels, lowering maintenance costs. 3. Affordable Hardware Compatibility: YOLOv10nano can run on cost-effective and inexpensive hardware platforms, keeping deployment costs low. This makes large-scale deployment feasible in commercial agricultural settings.

2.1.2. Augmentation

To create insect-free images that will serve as background in the image-augmentation process, we started by accessing the original image and removing all insects from it. We completed that by using the corresponding annotation file. To remove insects from the image, we utilized the inpaint function from the OpenCV library (cv.inpaint). The inpaint function is used for restoring (inpainting) areas of an image that are damaged or occluded. It helps “fill in” the damaged regions by leveraging information from the surrounding pixels. We chose the fast-marching method for this task because it was the fastest approach and produced results of sufficient quality during our initial tests. This method fills the damaged areas by starting at the boundaries and progressively extending the restoration inward, based on pixel priority. It uses the color information and structure of neighboring pixels to ensure both speed and basic continuity in textures.

We identified the areas containing insects using annotations from the Diopsis dataset. For these areas, we defined a mask with values of 255, and the function fills these regions with pixels from the neighboring areas. Initial tests with small insects showed promising results. However, we observed that for larger insects, the function sometimes leaves white spots, resembling a glare, on the image. We determined that this is not a significant issue. On the contrary, it adds complexity to the background image, resembling light reflections on the traps—a phenomenon that is quite common under real-world conditions.

The derived image constitutes a single canvas that we stored for image augmentation. We repeated the process, and we extracted 150 canvases that included all background variations that we visually assessed as the most important ones. For adding insects from other images onto these insect-free images, we followed a specific process that actually applies a mask onto the extracted images of insects so that they do not appear artificial in the synthesized image:

• Using the annotation file, we located the insects’ bounding boxes in the image.
• We created a copy of this selected area including the insect.
• The background within the bounding box was removed, turning it white.
• The copy was then converted to grayscale.
• We iterated through each pixel in the grayscale image, identifying pixels with values below 240 (with 255 representing pure white).
• For each identified pixel, we retrieved the corresponding pixel value from the original insect image (including all three RGB channels) and applied it to the corresponding location in the insect-free image.

This method is fast and allows us to seamlessly overlay insects onto background images, preserving the original insect details without artifacts.

2.1.3. Application of the SAHI Approach

SAHI, which stands for Slicing Aided Hyper Inference, is a method used in image recognition to improve the accuracy of object detection models, particularly for detecting small or densely packed objects [17]. It enables large images to be processed without the need to downscale them, preserving details that are crucial for accurate detection. The technique is often employed when using AI models, such as those based on deep learning, or in computer vision tasks where standard object detection models struggle to identify objects accurately in large or high-resolution images. SAHI divides the original large image into smaller, overlapping patches or slices. By processing these smaller slices instead of the entire image at once, the method enables more detailed examination of small objects, which would otherwise be overlooked or inaccurately detected in a full-resolution image. After slicing, we applied YOLO10 object detection models to each slice independently to detect objects within that smaller region. This approach increases the sensitivity of the model to detect objects that would appear tiny in the context of the full image. The results from each slice are then stitched back together to form a complete detection map for the entire image. After inference, SAHI aggregates the results from all slices, merging overlapping detections and ensuring that objects detected across multiple slices are only counted once. In images of insects, small insects might be hard to detect in the full image, but SAHI can improve the detection accuracy by analyzing smaller sections of the image in detail helping to detect the smaller insects or debris that would typically be missed in large images. If the sliced dimensions do not match YOLO’s input size exactly, the resizing step can be handled by the SAHI framework itself or within the inference pipeline.

In the annotations from the Diopsis dataset, there is no separate class for debris. The dataset includes only one class: insects. The algorithm learns to identify insects but does not distinguish between other objects that might appear in the image. As a result, debris that does not exhibit insect-like characteristics is simply ignored by the algorithm, just like any other non-insect object. Additionally, we set the algorithm’s confidence threshold to 0.25, which is relatively low. However, based on our tests, this threshold performs well, minimizing false positives without introducing significant errors. We did not observe any major issues with the algorithm during visual inspection. Specifically, we did not find instances where debris was mistakenly identified as insects to a degree that would cause problems. On the contrary, the algorithm successfully detects insects that are not explicitly annotated in the dataset. In the samples we reviewed, we often had to zoom into the image to confirm that the objects detected were indeed insects.

2.1.4. First Dataset

At first, we used the dataset as it was, in its original size without any augmentation or preprocessing other than what YOLO10 applies internally.

In Figure 2, we see that the insects can appear very small and with shapes that may be difficult sometimes to discern from debris.

2.1.5. Second Dataset

The second set included all images from the first set, along with an additional 2250 synthetic images to enrich the dataset. This brings the total number of images in the second set to 5164, each accompanied by its corresponding annotation file. The synthetic images were created through the following process:

Step 1: We selected 150 random images from the first dataset and removed the insects from each image, using the labeling file to locate each insect precisely. After removing the insects, we smoothed the gaps left in the background, creating 150 insect-free images (the so-called ‘backgrounds’ or ’canvases’).

Step 2: Next, we chose 15 random images from the first dataset that were not included in the 150 from Step 1. For each of these 15 images, we used the labeling files to identify the position of each insect, then copied and pasted the insect bodies onto each of the 150 background images from Step 1. This process yielded 150 × 15 = 2250 new synthetic images.

Finally, we organized the second set into folders for 5-fold cross-validation and converted the image highlights to YOLO format, consistent with the first set. This resulted in five folders of images, ready for validation and model training under the YOLO framework.

This procedure is illustrated in Figure 3.

2.1.6. Third Dataset

The third dataset is even richer than the second. During training with YOLOv10, we observed some loss in detection accuracy, particularly for small insects, due to resizing the original images (3280 × 2464 pixels) to 640 × 640 pixels. SAHI segmentation crops each image into pieces of 640 × 640 pixels, which is the same as the input size of YOLO. We thought that this size would be an image size suitable for distinguishing small insects without having to have a large fragmentation of the original images.

The new dataset that was made by dividing the original images into 640 × 640 segments resulted in 541,650 images. This approach aimed to improve SAHI efficiency by enabling the model to learn from smaller, context-specific images. Nevertheless, the results remained unsatisfactory, as the model lost contextual understanding of the full image and was challenging to train due to the sheer volume of data. We then enriched our dataset by incorporating segments from existing images at various scales, allowing the algorithm to learn from images resized differently from the original. The third dataset includes all images from the second dataset plus an additional 4077 randomly selected image sections, sized 480 × 480, 640 × 640, 720 × 720, and 1280 × 1280 pixels. This final dataset contains 5164 images at 3280 × 2464 pixels and 4077 images of varied dimensions, optimized to allow the model to learn from differently scaled images while ultimately training on 640 × 640 images (see Figure 4).

2.2. Hardware

The final device we used is a Raspberry Pi Zero w 2. This device has a 1 GHz quad-core 64-bit Arm Cortex-A53 CPU that has 512 MB SDRAM RAM and enables connection to Wifi 2.4 GHz 802.11 b/g/n wireless LAN. It has a low consumption of 125 mA at idle and up to 580 mA at full power. This device has one of the lowest consumptions of the Raspberry Pi family of devices and can run 64-bit operating system. As of 6 November 2024, its cost is approximately EUR 15–20, while the camera costs around EUR 20.

2.3. Object Recognition Algorithm

We decided to experiment with the YOLOv10nano [16] algorithm expecting images with a size of 640 × 640 pixels. We chose this algorithm as it is fast and is known to achieve good performance in recognizing small-sized objects such as insects in images. Another reason was to embed the smallest version of the algorithm, which is YOLOv10 nano, to a small full-computing board in the series of Raspberry Pi. This model has 2.3 million parameters, the smallest of all previous YOLO algorithms. It has good accuracy and a very short latency in the Coco dataset. Due to the limited resources of our device and the special characteristics of YOLOv10 nano, we felt that it is the most suitable model to allow for an embeddable and effective solution for further experiments.

By choosing such a small model, we knew from the outset that its performance would be compared to the larger models, especially with the large size of the images (resolution 3280 × 2464 pixels) from the dataset. The goal is to have the device with the lowest consumption (and therefore optimize the limited resources), resulting in the lowest cost with an acceptable measurement accuracy.

3. Results

This study’s detection models and image processing methods were evaluated using three distinct datasets based on the Diopsis dataset. The results are gathered in

Table 1. Five-fold evaluation on different datasets. All YOLO versions in experiments A–G, with 640 × 640 images.

	mAP50 (%)
Fold	A	B	C	D	E	F	G
1	63.9	33.6	66.7	28.0	67.3	57.7	71.8
2	63.6	33.6	70.1	34.3	67.0	55.5	72.7
3	64.8	39.5	66.7	32.9	69.8	55.4	73.2
4	62.7	35.8	66.0	36.9	67.1	54.7	72.5
5	62.8	38.0	67.4	33.2	67.8	57.5	73.5
mean	63.6	36.1	67.4	33.1	67.8	56.2	72.7
std	0.9	2.6	1.6	3.2	1.2	1.4	0.7

A: The Diopsis data in their original size (3280 × 2464 pixels). The YOLO will resize the images to 640 × 640. B: The second dataset. The algorithms have been trained in the original dataset on 3280 × 2464-pixel images like in A, but the algorithms have been executed on the test set with the SAHI method by hashing the images into 640 × 640 pixels. C: The A dataset with the additional images of the image augmentation process (insect relocation onto various canvases). D: The second dataset (SAHI processing). The same approach as in C, but with the SAHI method by hashing the images into 640 × 640 pixels. E: The C dataset with the third dataset on images with mixed sizes of 3280 × 2464, 480 × 480, 640 × 640, 720 × 720, and 1280 × 1280 pixel crops. F: This is the E dataset with SAHI applied. The results of the algorithms in the test are set in 3280 × 2464 pixel images. The algorithms have been trained in the third dataset on images with mixed sizes of 3280 × 2464, 480 × 480, 640 × 640, 720 × 720, and 1280 × 1280 pixels, but the algorithms have been executed in the test set using the SAHI method by fragmenting the images into 640 × 640 pixels. G: The datasets E and F with non-maximum suppression applied. The algorithms have been trained in the third dataset on images with mixed sizes of 3280 × 2464, 480 × 480, 640 × 640, 720 × 720, and 1280 × 1280 pixels, but each result is the result of two algorithms, the first being simple YOLOv10 and the second being YOLOv10 executed with the SAHI method with fragmentation of the images into 800 × 800 pixels. In the combined pool of bounding boxes, we applied non-maximum suppression to keep the different annotation boxes.

and are based on the mAP50 metric. The mAP50 measures the model’s accuracy by calculating the average precision across all classes at a specific Intersection over Union (IoU) threshold of 0.5. IoU measures the overlap between the predicted and ground-truth bounding boxes. A higher mAP50 indicates better model performance in detecting objects with at least 50% overlap. A stricter overlap is not necessary as human labelling of insects in images is not very precise.

1.

Performance Metrics Across Datasets

First Dataset: Original Diopsis dataset without augmentation or SAHI processing. The average mean Average Precision (mAP50) achieved was 63.6% for insect detection.

Second Dataset: Original dataset with 2250 synthetic images created through background extraction and insect repositioning. Training on this augmented dataset yielded a slight improvement, with a mAP average of 67.8%.

Third Dataset: The most extensive dataset, combining the second dataset with additional images of varied scales (480 × 480, 640 × 640, 720 × 720, and 1280 × 1280 pixels), bringing the total image count to 9241. The mAP50 score further improved to 72.7%, particularly when applying SAHI processing.

2.

SAHI Processing Impact

The application of SAHI processing significantly impacted detection rates across the datasets:

Non-SAHI vs. SAHI: For the second dataset, non-SAHI processing yielded a mean mAP50 of 67.8%, whereas SAHI processing increased it to 71.8%.

Effect on Small Objects: SAHI processing demonstrated its effectiveness in detecting smaller insects by dividing high-resolution images into smaller, overlapping sections. This approach mitigated image scaling issues but required additional processing time.

Summary of Key Metrics:

First Dataset (Original): YOLO10n on the original Diopsis images and annotation files, mAP50 of 63.6%, and avg. processing time of 5.15 s.

Second Dataset (Synthetic Augmentation): mAP50 of 67.8% and avg. processing time of 76.4 s.

Third Dataset (SAHI + Mixed Scale): mAP50 of 71.8% and avg. processing time of 87.3 s.

In summary, augmenting the dataset through background extraction and utilizing SAHI processing contributed significantly to enhancing detection performance across all metrics, albeit with computational trade-offs that may inform future device-specific optimizations.

Results on Device:

The device used for this study is the Raspberry Pi Zero W2. The model that was implemented was trained using the original Diopsis dataset.

Our initial method achieved an average processing time of approximately 5 s, with an mAP50 of 67.3%. This processing time is deemed acceptable, with the algorithm requiring minimal resources, resulting in low energy consumption for the device. The predictive accuracy is considered to be adequate given the limitations previously discussed.

The third method yielded the highest mAP50 score of 71.8%, but with a significantly increased processing time of approximately 1 min and 27 s. This longer runtime causes substantial energy consumption and delays. However, if insect identification occurs once or twice daily, this method remains viable. Furthermore, when integrated with a microcontroller-based on/off mechanism for the Raspberry Pi within the trap, as outlined in previous research, the third method could effectively extend the operational lifespan of the trap, despite its higher energy demands. In

Table 2. Three different approaches running on Raspberry Pi W2. Mean processing time of an image in sec.

Approaches	mAP50 (%)	Time (s)
YOLOv10n	67.3	5.15
SAHI 640 × 640	57.7	103.34
Combine YOLOv10n+SAHI 800 × 800	71.8	76.37

, we gather the results on embedding different approaches in an edge device. The results in Table 1 have been produced using a PC with a GPU and the results in Table 2 using a Raspberry Pi Zero W2.

Initial Processing: Using the first dataset and a standard YOLOv10 configuration, the average detection processing time on the Raspberry Pi Zero W 2 was approximately 5.15 s per image, achieving an acceptable mAP50 of 67.3%.

In [18], we present figures comparing the detection results from different approaches (e.g., before and after augmentation, SAHI vs. YOLOv10nano). These figures are not included here, as resizing the images reduces the visibility of insects and their annotation boxes.

4. Discussion

This study presents an innovative approach to insect detection by focusing on background extraction and augmentation, rather than solely on the insect itself. Through the use of the Diopsis dataset and various augmentation techniques, we generated synthetic datasets that enhanced model robustness to environmental variability, such as the shadows, overlapping objects, and uneven lighting conditions that are commonly encountered in field setups. Specifically, the augmentation process involving background isolation and insect repositioning on these insect-free canvases demonstrated a marked improvement in detection consistency, particularly when deployed on resource-limited hardware like the Raspberry Pi Zero W2. The application of Slicing Aided Hyper Inference (SAHI) further enhanced detection accuracy for small and densely packed insects by allowing for a detailed analysis of high-resolution images, despite the constraints imposed by the hardware. Note that in recent years, new data augmentation methods, such as Generative Adversarial Networks (GANs) [19], have emerged. These models learn to generate new data with similar statistical properties to the training set. Their potential role in augmenting insect figures remains an open research area.

Table 2 shows that SAHI cannot be beneficial in all circumstances, and we found an indirect way to make use of it by increasing map accuracy from 67.8 to 72.7%. The extraction of backgrounds and their use as an augmentation canvas increased the mAP50 of the first dataset from 63.6% to 67.4% in the second dataset. We then combined SAHI and image augmented dataset, and we reached the final mAP50 of 72.7%.

Although using SAHI and larger datasets improved model performance, particularly in terms of the mAP50, this approach required a trade-off in processing time. The SAHI method, while boosting the detection rates, resulted in significant increases in computational time, which raises concerns for scenarios requiring rapid processing. Nevertheless, for applications with lower frequency demands, such as daily or semi-daily monitoring, the benefits of this methodology in terms of accuracy may outweigh the delays introduced by the SAHI processing and additional augmentation.

We analyzed briefly the causes of failure in the counting process. In dataset A, where YOLOv10nano was trained solely on the original dataset, YOLOv10nano misses detection because resizing to the dimension that YOLO expects makes small insects invisible. A general issue with the Diopsis dataset that affects our experiments from A to G is the incomplete annotation of many images. However, our models manage to detect even the unannotated insects, demonstrating their effectiveness. Nevertheless, since we used the official annotations from the Diopsis dataset for evaluation, this affected our reported performance metrics negatively. Finally, our approach sometimes registered false positives where an insect receives multiple bounding boxes. This issue stems from non-maximum suppression (NMS) and can be resolved by adjusting the IoU threshold.

5. Conclusions

In this work, we introduced a novel background-focused approach to augmenting insect detection data, effectively enhancing model performance on low-power edge devices under field conditions. In this work, the background is in focus and not the insects themselves. The augmentation strategy proved effective in reducing variability due to environmental factors, and the SAHI methodology enabled the detailed detection of small objects within large images. Although trade-offs were observed in processing time, the implemented methodology offers a practical and scalable solution for insect monitoring in real-world, constrained environments. Future work may explore optimizing SAHI processing to reduce computational demands, as well as further testing across different insect populations and environments to validate the approach’s generalizability and efficacy in diverse ecological monitoring applications.