Detection of Mosquito Breeding Areas using Semantic Segmentation

Abstract

The combination of deep learning (DL) and convolutional neural networks (CNN) with image analysis to locate stagnant water will play a crucial role in the dengue control process. This paper aims to automatically segment stagnant water areas in aerial images, acquired by a drone camera, using the latest CNN semantic segmentation method (SegNet). To enhance the effectiveness of our system and as the solution for the lack of dataset, we utilise two different datasets with high domain feature correlation. In our project, pre-training is first done on a large generalised dataset with areas of water, and then the trained model with trained weights is retrained using a task-specific dataset. It should be noted that the conditional distribution of the labels is similar for both datasets. The performance of the SegNet was evaluated with pixel accuracy and dice score. The model exhibited an accuracy of 80% and a dice score of 91%, indicating that our proposed method is efficient to segment water in RGB aerial imagery.