Shoreline Recognition Using Machine Learning Techniques

Abstract

Coastal areas have emerged to be the most significant and dynamic regions worldwide. Therefore, automating shoreline recognition will aid non-profit conservation authorities to reduce public budget expenditures, relieve erosion damage, and increase the climate resilience of the natural environment. In this paper, advanced ML boosting algorithms including XGBoost, and LGBM are firstly applied into shoreline recognition with aerial images (of Lake Ontario in this study). This paper first discussed the significance and a literature review of recent progress in shoreline detection. Then, this paper adopted semantic segmentation instead of detecting shoreline directly, which enables the (Machine Learning) ML model to achieve relatively high accuracy with a small amount of data. 5 high-resolution images are used for training the model in which shorelines are detected. The work was carried out in four steps: 1) labeling the contents of shoreline images as areas of water and banks; 2) training ML algorithms; 3) using the trained algorithms to classify the image content as either water or land objects; 4) post-processing by de-noising image pixels (applying a Fourier transform algorithm) to obtain a defined shoreline. The averaged training time per image for Random Forest, XGBoost, and LGBM algorithms are 195.2 sec, 71.0 sec, and 8.6 sec, respectively. The averaged accuracy is 95.6%, 96.0%, and 94.8%, respectively; the XGBoost algorithm has slightly higher accuracy, while LGBM has a significantly shorter runtime. Cross-validation of the LGBM algorithm reduced the training time by around 23% (7.0 sec) and increased the accuracy by only 1.1% (to 95.9%).