High-resolution UAV dataset for identification of breeding sites for malaria and dengue vectors in rural and peri-urban Tanzania: An experimental approach with Gemma 3N E4B architecture

Abstract

Malaria and Dengue are the major vector borne diseases (VBDs) with high prevalence in the tropics, where their existence is significantly linked to the changing climatic conditions which favour the continuous breeding of vectors. In Tanzania, the burden for malaria is more pronounced in rural and peri‑urban settings than in urban, with prevalence of up to 24% for under 5 children, varying across geographical zones. This paper presents a high-resolution UAV dataset collected from Mtwara and Tabora regions in southern and western zones, respectively, for development of surveillance models for potential mosquito breeding sites. The presented dataset contains 19 classes of mosquito breeding sites that were identified from 471 aerial images as a 80.5% of retained images with useful information from the total of 585 collected in Igunga and Masasi Districts in Tabora and Mtwara regions, respectively. With an experimental approach, the paper presents the use of the imagery dataset on the state of art Gemma 3 N E4B, an instruction-tuned multimodal architecture capable of processing both visual and textual data. By using the 80.5% proportion remained after collaborative image labelling, we trained and validated the instructional model with exceptional convergence characteristics and a stabilization phase yielding a loss of 0.0319. The reduction in loss was 96.7%, from 9.620 to 0.319, within 117 steps.

Specifications Table

Subject	Computer Sciences
Specific subject area	Object detection and segmentation tasks to identify potential breeding sites for vectors (Anopheles and Aedes mosquitos) in selected rural and peri urban settings
Type of data	Image
Data collection	Data were collected by using a Drone: Type: DJI Phantom 4 pro, flight height: 40MTRS from the ground, Camera angle: Nadir (90 Degrees Downward). The drone was connected to an android phone for quality check. The DJI GO 4 app was installed on an android phone and with a USB cable the mobile phone was connected to the remote controller and the drone. Mosquito breeding sites were the points of interest for the images taken which included household surroundings, agricultural land, tyre tracks, ponds, uncovered well among others. Total number of images collected was 585, and a participatory approach was used for the multi-class labelling where 471 images could be retained with useful information.
Data source location	Affiliation: •The Nelson Mandela African Institution of Science and Technology •Arusha •Tanzania The dataset was collected from the following locations: •Igunga District, Tabora and Masasi District, Mtwara •Tanzania
Data Accessibility	Repository name: Kaggle Data identification number: NA Direct URL to data: https://www.kaggle.com/datasets/devothanyambo/images-for-the-detection-of-mosquito-breeding
Related research article	NA

1. Value of the Data

• •To develop a deep learning model for the detection of breeding sites for malaria and dengue vectors (anopheles and aedes)
• •The dataset can be used to enhance models that can be used in limited resource settings and where the cost of acquiring local data is prohibitive. The images are labelled based on a multi-class approach to describe multiple features within images
• •The dataset integrates household, agricultural and natural features that may contribute to mosquito breeding, especially in African settings
• •Our dataset integrates nineteen (19) relevant conditions (classes) for mosquito breeding, where the chances of Aedes breeding (after the drying out of water holding sources) is captured at a high resolution. Through this, models can be developed to evaluate and warn against the risk of Aedes re-occurring at the onset of short rainy season.
• •To the best of the authors’ knowledge, this is the biggest high-resolution UAV imagery dataset collected from malaria endemic regions where Anopheles and Aedes are proven to co-exist.

2. Background

The goal of creating the imagery dataset is to strengthen the environmental assessment and the detection of mosquito breeding sites by community members as an approach of fighting against vector borne diseases such as malaria and dengue. Collection of the dataset was influenced by potential co-occurrence of malaria and dengue diseases in the tropics [1]. As such our focus was on high endemic areas in Tanzania (Kagera, Mtwara and Tabora) [2]. For locations where dengue disease is not prioritised, the breeding of anopheles and aedes is likely to affect the community members; leaving the burden of dengue unnoticed.

The dataset will be used by researchers in developing detection models for mosquito breeding sites for early intervention against VBDs. Moreover, the dataset can be used to develop models that can differentiate potential breeding sites for most disease vectors whose growth conditions are promoted by stagnant water, trash, uncontrolled grass cover, and intensive irrigation farms.

Growth differences between aedes and anopheles mosquitos require significant differentiation with respect to water holding surfaces because Aedes eggs are able to survive dry conditions after water drainage while Anopheles cannot. This difference requires images that capture drained settings to still be identified as they could be potential carriers of Aedes eggs, posing an infection risk all year round [3]. Our dataset integrates these settings in labels and instruction that is added in the images example conversations as outlined in the sample (Appendix I).

3. Data Description

The labelled dataset represents a critical resource for advancing deep learning applications targeting vector-borne diseases, particularly malaria and dengue. These diseases, transmitted by Anopheles and Aedes mosquitoes respectively, continue to affect millions globally, with sub-Saharan Africa bearing a significant burden. Despite advances in malaria vector surveillance and control in Tanzania and globally, existing datasets remain limited in scale, resolution, and contextual relevance to African environments [4,5]. Our dataset directly addresses this gap by offering high-resolution drone imagery annotated for multiple environmental conditions associated with mosquito breeding.

3.1. Image collection and processing

A data set of 585 high-resolution aerial images was collected by NM-AIST and labelled in collaboration with experts from the National Institute for Medical Research (NIMR), Tanzania. We used a 20-megapixel camera mounted on a DJI Phantom 4 Pro UAV, flying at 40 m above ground. Study sites were in Mtwara (Masasi District: Mkomaido, Mwengemtapika, Roben villages) and Tabora (Igunga District: Simbo village), chosen for their high malaria prevalence (20 % and 23 % in each region, respectively. The areas also reflect diverse ecological and anthropogenic environments, including household compounds, rice paddies, drainage systems, natural water collections, and agricultural fields. After manual inspection and labeling, 471 images (80.5 % of the original 585) containing useful information were retained (see per-class counts in Table 2). The selected images reflect diverse rural and peri‑urban settings (e.g., brick and mud houses, trash, plastic waste) which we ensured by sampling across different village environments.

Table 2.

dataset distribution for training, validation and testing.

Category	Number of Images	Percentage ( %)
Training	377	80
Validation	47	10
Testing	47	10

3.2. Class structure and labelling approach

The dataset includes annotations for nineteen (19) unique classes relevant to mosquito breeding ecology. These classes span both natural and artificial features such as temporary pools, stagnant drains, animal troughs, banana plantations, discarded containers, many of which are rarely included in existing imagery datasets for VBDs detection. The annotations are multi-class and spatially descriptive, allowing models not only to detect a breeding condition but to interpret contextual co-occurrence (e.g., proximity to households or shaded environments). This adds granularity and improves the models’ ability to generalize to varied terrain and scenarios, especially in low-resource or hard-to-monitor regions.

3.3. Addressing dengue vector breeding conditions

While malaria control has historically received more attention, the breeding conditions for Aedes aegypti, the dengue vector, are often overlooked in surveillance datasets. Aedes species prefer temporary, clean water habitats often found in urban or peri‑urban containers, flower pots, and tree holes that may dry out between rains. Our dataset explicitly captures high-resolution imagery of such post-drying habitats, enabling the modelling of re-emergence risks particularly at the onset of short rainy seasons, a time when Aedes proliferation can spike rapidly (Fig.1). On the other hand, we show the continuous supportive breeding sites for Anopheles through stagnant water in temporal pools and agricultural settings (Fig. 2). Neglecting these ephemeral breeding habitats in vector monitoring poses significant consequences as it limits the predictive power of early warning systems for dengue outbreaks, underestimates transmission risk during inter-epidemic periods, and leaves health systems underprepared. By documenting these features at high resolution, our dataset uniquely enables AI systems to forecast potential re-colonization zones even when visible water is absent.

Fig. 1.

High resolution satellite images showing vegetation cover in a Masasi Mtwara, peri‑urban household surroundings with potential dried water holding surfaces conducive for Aedes.

Fig. 2.

High resolution satellite images showing stagnant water bodies around households and agricultural lands, prominent conditions for the growth of Anopheles.

4. Experimental Design, Materials and Methods

4.1. Field data collection

A 585 images dataset was collected in a research project by the Nelson Mandela African Institution of Science and Technology (NM-AIST) and labelled in collaboration with experts from the National Institute for Medical Research (NIMR), Tanzania. A 20 MP camera was used mounted on a DJI Phantom 4 pro UAV. Mtwara and Tabora were the regions of study, selected due to their overall malaria endemicity rate, 20 % and 23 %, respectively, according to the 2022 – 2023 Tanzania Health Demographic Survey, prevalence of malaria in children by region [2]. Mkomaido, Mwengemtapika, and Roben villages for Masasi district in the Mtwara region were included, while in the Tabora region, in the Igunga district, Simbo village was included. The villages selection considered a combination of agriculture-intense and commercial orientation of the population. The UAV was set at an altitude of 40Mtrs above the ground ensuring a 20MP resolution rate for all the images collected.

Community members were not given prior information to avoid potential emergency cleaning that would have diluted the real conditions on the ground. Images were captured for four days between 4th – 12th April 2025 in the two regions. For both regions, this was a heavy rains season and it was adequate for capturing all possible water holding surfaces. The images were captured by researchers and labelled in collaboration with domain experts from NIMR after a first round on creating initial instruction-label pairs for the multimodal dataset by using a large language model. The LLM was prompted with task-specific templates covering classification, reasoning, and vision-language grounding. All LLM generated labels were subsequently reviewed by human annotators with domain expertise to correct inaccurate labels, removal of low-quality samples, and adherence to task definitions. Upon labelling, 80.5 % of the images were retained for use with the multi-class labels. Nineteen (19) classes were identified and their respective count in the dataset is indicated in Table 1 where 471 images were successfully labelled. Table 1 shows a total of 676 instances as multi-class labels for the 471 images; as such majority of the images contained more than one class. The geographical diversity of the study sites ensured the capture of environments with brick houses in rural and peri urban settings, mud with grass roof houses, houses with and without mesh windows, uncontrolled trashing and plastic waste.

Table 1.

Per-class image counts.

Rank	Label	Images
1	no_breeding_site	265
2	temporary_pool	184
3	rice_paddies	84
4	tyre_track	38
5	solid_waste	27
6	foot_print	22
7	open_container	19
8	roadside_ditches	9
9	open_septic_tank	8
10	borehole	6
11	abandoned_building	4
12	tomato_farm	2
13	fish_pond	2
14	thatched_house	1
15	temporary_house_foundation	1
16	maize_farm	1
17	mud_house	1
18	abandoned_toilet	1
19	solid_waste_pit	1

Images were captured for four days between 4th – 12th April 2025 in the two regions. For both regions, this was a heavy rains season and it was adequate for capturing all possible water holding surfaces. The geographical diversity of the study sites ensured the capture of environments with brick houses in rural and peri urban settings, mud with grass roof houses, houses with and without mesh windows, uncontrolled trashing and plastic waste.

4.2. Experimental modelling

An experimental modelling was done to evaluate the dataset usability. The aims were to develop and evaluate a baseline multimodal AI model capable of identifying mosquito breeding sites from aerial drone imagery and generating context-aware recommendations to mitigate vector proliferation. The experimental modelling was informed with a logical flowchart of the study steps from field data collection to the validation of an instructional model for the detection and interpretation of contextual co-occurrence of mosquito breeding cites (Fig. 3).

Fig. 3.

Flow chart showing steps of the study from data collection to the testing of Gemma 3 N E4B Architecture for the detection of mosquito breeding sites.

The dataset was split following the 80/10/10 for training, validation and testing respectively as shown in Table 2 below.

The model used was Gemma 3 N E4B (https://huggingface.co/google/gemma-3n-E4B) an instruction-tuned multimodal architecture capable of processing both visual and textual data. The base model incorporated a sequence length of 2048 tokens and utilized 4-bit quantization for efficient memory usage. Fine-tuning was performed using the Low-Rank Adaptation (LoRA) technique to ensure efficiency based on introduction of new knowledge on top of the original models [6].

Key Configuration Parameters:

• •LoRA Rank: 16
• •Alpha: 16
• •Dropout: 0.05
• •Trainable Parameters: 38.4 million (0.49% of total model size)

The training dataset consisted of 80% of the labelled imagery samples; each paired with structured multi-turn dialogues simulating interactions between a user and an AI assistant. Each entry included:

• •An image (DJI JPG format)
• •A visual description
• •Annotated breeding site types and descriptions
• •Mitigation recommendations
• •Follow-up questions

Appendix I shows an example of one entry into the training process.

4.3. Training environment

The model was trained on a Tesla T4 GPU with 14.741 GB of RAM, running on a Linux platform. The training environment incorporated PyTorch version 2.6.0, Transformers 4.53.3, and Unsloth 2025.7.8 for optimized performance. An 8-bit AdamW optimizer with cosine learning rate scheduling was used to enhance efficiency. The training was conducted over a single epoch with a batch size of 1, supported by gradient accumulation steps of 4. Mixed precision (FP16) was enabled to improve memory and computational efficiency, along with checkpointing to manage training progress and resource usage effectively.

4.4. Optimization strategy

Two-stage optimization was done:

• a)
  Core Training Configuration

  • •
    Epochs: 1 (which was sufficient for specialized domain adaptation)
  • •
    Batch Size: 1 per device
  • •
    Gradient Accumulation: 4 steps (effective batch size: 4)
  • •
    Learning Rate: 1e-4 with cosine scheduling
  • •
    Warmup Steps: 10 (gradual learning rate increase)
  • •
    Weight Decay: 0.01 (L2 regularization)
• b)
  Advanced Optimization Techniques

  • •
    Mixed Precision: FP16 for memory efficiency and speed
  • •
    Gradient Checkpointing: was enabled for memory optimization
  • •
    Smart Gradient Offloading: Unsloth-specific VRAM management
  • •
    Response-Only Training: Focused learning on assistant responses
  • •
    8-bit Optimization: AdamW optimizer with reduced memory footprint

4.5. Experiments

• Loss Convergence Profile

The training demonstrated exceptional convergence characteristics with stabilization phase yielding a loss of 0.0319, a reduction by 96.7 %, from 9.620 to 0.319, within 117 steps (Table 3).

Table 3.

Model convergence characteristics.

Training Phase	Loss Range
Initial (Steps 1–20)	9.620 → 6.786
Steep Descent (Steps 20–30)	6.786 → 0.954
Refinement (Steps 30–75)	0.954 → 0.020
Stabilization (Steps 75–117)	0.020 → 0.319

4.6. Key performance metrics

Expert validation was done on the model outputs especially relating it to the ground truth. The example of dataset conversation structure given in Appendix I highlight the strengths of the optimized model. Objective metrics considered were loss reduction, convergence quality, training stability and loss, as described below:

• •Loss Reduction: 96.7 % improvement (from 9.620 to 0.319)
• •Convergence Quality: Smooth, monotonic improvement without overfitting
• •Training Stability: Minimal loss oscillation in final phases
• •Optimal Performance: Lowest loss of 0.019 achieved at step 75

4.7. Resource utilization and efficiency

Memory Management on training:

• •Peak Memory Usage: 12.592 GB (85.4 % of available GPU memory)
• •Training Memory Overhead: 0.0 GB additional (100 % efficiency)
• •Memory Optimization: Unsloth smart gradient offloading enabled
• •Memory Efficiency: Optimal utilization without out-of-memory conditions

Training Duration and Throughput:

• •Total Training Time: 15.05 min (903.18 s)
• •Training Speed: approximately 8 examples/minute
• •Computational Efficiency: 2 x speedup achieved through Unsloth optimization approach
• •GPU Utilization: Consistently high throughout training process

4.8. Model validation and testing

Remained 20 % of the labelled dataset was used for validation (10 %) and testing (10 %) to confirm the models’ specialization on tasks and knowledge retention.

(a) Specialized Task Performance

Post-training validation showed a successful domain adaptation on the following aspects:

• (i)Mosquito Breeding Site Analysis: Model demonstrates accurate identification of potential breeding environments including rice paddies, standing water features, and drainage systems
• (ii)Contextual Understanding: Proper interpretation of aerial imagery with relevant environmental analysis
• (iii)Recommendation Generation: Appropriate mitigation strategies based on identified risk factors

(b) Knowledge Retention Assessment

Testing confirmed preservation of general knowledge capabilities on three domains:

• (i)Factual Knowledge: Maintained accuracy on general world knowledge
• (ii)Reasoning Abilities: Preserved logical thinking and explanation capabilities
• (iii)Language Fluency: Retained natural language generation quality as validated by human experts.

The main objective of the imagery dataset is to create an instructional tool that can be used by community health workers and members to identify and learn about mosquito breeding sites around their locations. The re-training process for the Gemma 3 N E4B architecture evaluated model’s specialization, suggestions on how to deal with identified breeding sites and general knowledge composition which will aid on overall advisory process at community levels.

• •Specialization: Accurate detection of temporary pools, rice paddies, roof tops, tyre tracks, and drainage systems among others as breeding sites
• •Mitigation Suggestions: Effective and context-aware, e.g., draining, larviciding, soil-filling
• •General Knowledge Retention: Model preserved prior reasoning and language fluency both in Kiswahili and English.

Limitations

This study appraises as practice on battling vector borne diseases by utilising aerial satellite imaging to eliminate breeding sites as well supported by [7]. The presented dataset is limited to generic environmental without breeding sites identifiable from specific household conditions such as small flower pots that could not be picked from 90 degrees downward angle. This limitation was ignored because of the potential to complement existing dataset that reflect household conditions such as those reported by [8] among others. The dataset is also bound on identifying potential breeding sites and not confirmed ones in order to promote early actions by the communities. For the identification of mosquito breeds and abundance by using walkthroughs, our dataset is limited since not all collected sites could be characterised in the current study [9].

Training and optimization limitations include the collected clean imagery dataset size (471 samples – 80 % made available after collaborative labelling), which may constrain the model’s generalizability to unseen landscapes [10]. However, the structure and quality of the dataset (multi-turn dialogues and contextual descriptions) enhanced learning efficiency. Future iterations could explore augmenting the dataset with additional household environmental variations and increasing interaction complexity.

Ethics Statement

The authors have read and follow the ethical requirements for publication in Data in Brief and confirming that the current work does not involve human subjects, animal experiments, or any data collected from social media platforms.

CRediT Author Statement

Devotha G. Nyambo: Conceptualization, Methodology, Software, Validation, Analysis, Investigation, Writing, Project administration, Funding acquisition.

Tumaini M. Lyimo: Methodology, Software, Images labelling, Model training.

Elingarami Sauli: Conceptualization, Validation, Review, Project administration, Funding acquisition.

Data Availability

• KaggleImages for the detection of mosquito breeding (Original data).