Detection and classification of unwanted plants and plastics present in aquatic bodies for recycling

PurposeOne of the major challenges faced by the world at present is management and treatment of waste. Especially, waste such as polyethylene (plastics) is non-degradable and is causing great damage to our environment. Aquatic environment is one among them that is getting affected by these plastic wastes. Water pollution is a great issue faced in many countries and steps to reduce it are being taken on a wide scale. Unwanted aquatic plants grown in ponds and lakes create problems like totally covering up the surface of the lake that blocks the sunlight for aquatic species and also reducing their total storage. Identifying such unwanted plants and plastics is a very essential part in treating and management of waste. Detection and classification help us to achieve this. With the help of satellites, drone-shot images of many oceans are captured, and the amount of plastic content present is detected using artificial intelligence. In artificial intelligence, we have many algorithms and platforms that help us to achieve object detection. Tensorflow is one such framework that helps us to perform object detection with the help of pre-trained models present in it, and thus, it is used in this study. Object detection uses computer vision to detect objects from images. Convolutional neural networks are a subset of machine learning that is helpful in image processing – in other words, processing of pixel data. In this study, we used the ResNet-50 model involving transfer learning for classifying unwanted plants and plastics. Lakes and ponds are the major places among the other aquatic environments where these kinds of wastes are found, and therefore, this study concentrates on waste present in these aquatic bodies. The lakes and ponds present near residential areas act as a place for storing excess rainwater, which prevents flooding. Many cities, especially residential areas, face a lot of water stagnation problems during the rainy season. Ponds and lakes near these areas contain unwanted plants and plastics present, which makes it a problem to store the rainwater that comes during monsoon. Another problem is that they don’t provide sunlight to enter deep into water, making the aquatic species difficult to survive. Preserving and maintaining such lakes from getting filled with non-degradable plastics and unwanted plant growth becomes very important. Therefore, the lakes and ponds present in such residential areas would be useful to detect the unwanted waste.Design/methodology/approachIn this study, the focus is on detection and classification of the plastics and unwanted plants. The dataset is very important for this study, which is an image dataset. There was not any readily available image data of unwanted plastics available online, and therefore, the images were captured from the lakes and ponds in Kanchipuram district. Images of duckweed, plastics, bulrush and leaves of sky lotus were taken. This dataset consisted a total of 200 images, with 50 images belonging to each category. Having this as the dataset, detection and classification were carried out.FindingsThe object detection took place for the plastic, duckweed, bulrush and leaves of sky lotus and the performance metrics such as precision and recall was evaluated to test the accuracy of the detections. Precision is used to calculate the number of correctly identified positive identifications. This is done by dividing the sum of true positives and false positives from the number of true positives. True positives are nothing but the number of correct predictions of positive identifications, and false positives are the number of false predictions of positive identifications. Similarly, recall is used to calculate the number of actual positives identified. We can calculate recall by dividing the sum of true positives and false negatives from the total number of true positives. Here false negatives are the number of false predictions of false identification. This performance metrics was evaluated for the trained model, and we obtained an average precision of 0.81 and an average recall of 0.86. The high precision and recall values of our model show that the model produces accurate results. Therefore, the model is producing good performance in detecting the unwanted plants and plastics from lakes and ponds. The evaluation results were visualized with the help of TensorBoard and are available in fig-4 and fig-5. The loss rate is visualized and is available in fig-6. We can see that the loss rate has reduced over the steps as we pass from 1,000 to 4000th step.Originality/valueThe work was originally carried out in the Kanchipuram district of Tamil Nadu.