Development of a Convolutional Neural Network Image Detection Model to Identify the Trophic Status of Lakes

Abstract

The eutrophication of lakes is a major concern in Bangladesh since it is largely responsible for water quality degradation in lakes throughout the country. As per previous research, lake trophic status may be identified using different methods. This study was conducted to seek a new method for determining trophic status across Dhaka City. A total of 110 lake water samples were collected from Banani, Gulshan, Hatirjheel, Ramna, Dhanmondi, Uttara, Nirjhor, Shaheed Sharani I & II, and Dhaka Cantonment Lakes, respectively. Lake surface images were captured at different angles, as part of the sampling process. The samples were tested for nitrate, phosphate, and chlorophyll-a concentration. These parameters were used to compute a trophic state index for each sample. The results indicated that 55% of the lake samples were hypereutrophic, 25% were eutrophic, 11% were mesotrophic, and only 9% were oligotrophic. The indices were used to label the corresponding lake images into the four categories. The images were divided into a training batch and a test batch. A Convolutional Neural Network (CNN) image detection model was designed on a Kaggle platform using Python and fed the images from the training batch. After the training, the model was tested using the other batch of images and displayed an overall accuracy of 79.4% in identifying the trophic states of the lake samples based on their images. The image detection method thus demonstrated higher efficiency than manual computation methods of identifying trophic status.

Journal of Engineering Science 16(1), 2025, 01-10