Mosquito-borne diseases have emerged in North Borneo in Malaysia due to rapid changes in the forest landscape, and mosquito surveillance is key to understanding disease transmission. However, surveillance programmes involving sampling and taxonomic identification require well-trained personnel, are time-consuming and labour-intensive. In this study, we aim to use a deep leaning model (DL) to develop an application capable of automatically detecting mosquito vectors collected from urban and suburban areas in North Borneo, Malaysia. Specifically, a DL model called MobileNetV2 was developed using a total of 4880 images of Aedes aegypti, Aedes albopictus and Culex quinquefasciatus mosquitoes, which are widely distributed in Malaysia. More importantly, the model was deployed as an application that can be used in the field. The model was fine-tuned with hyperparameters of learning rate 0.0001, 0.0005, 0.001, 0.01 and the performance of the model was tested for accuracy, precision, recall and F1 score. Inference time was also considered during development to assess the feasibility of the model as an app in the real world. The model showed an accuracy of at least 97%, a precision of 96% and a recall of 97% on the test set. When used as an app in the field to detect mosquitoes with the elements of different background environments, the model was able to achieve an accuracy of 76% with an inference time of 47.33 ms. Our result demonstrates the practicality of computer vision and DL in the real world of vector and pest surveillance programmes. In the future, more image data and robust DL architecture can be explored to improve the prediction result.

由于森林景观的快速变化，马来西亚北婆罗洲出现了蚊媒疾病，而蚊子监测是了解疾病传播的关键。然而，涉及采样和分类鉴定的监测计划需要训练有素的人员，既耗时又费力。在这项研究中，我们的目标是使用深度学习模型 (DL) 开发一款能够自动检测从马来西亚北婆罗洲城市和郊区收集的蚊媒的应用程序。具体来说，使用总共 4880 张广泛分布于马来西亚的埃及伊蚊、白纹伊蚊和致倦库蚊图像开发了一个名为 MobileNetV2 的深度学习模型。更重要的是，该模型被部署为可在现场使用的应用程序。使用学习率0.0001、0.0005、0.001、0.01的超参数对模型进行微调，并测试模型的准确率、精确率、召回率和F1分数的性能。在开发过程中还考虑了推理时间，以评估模型作为现实世界中的应用程序的可行性。该模型在测试集上的准确率至少为 97%，精确度为 96%，召回率为 97%。当作为应用程序在现场使用不同背景环境的元素检测蚊子时，该模型能够达到 76% 的准确率，推理时间为 47.33 毫秒。我们的结果证明了计算机视觉和深度学习在病媒和害虫监测项目的现实世界中的实用性。未来，可以探索更多的图像数据和鲁棒的深度学习架构来改善预测结果。