Image Based Ecological Instream Water Monitoring Models
The CT DEEP Water Monitoring and Assessment Program (MAP) has developed and implemented a low-cost and innovative approach to evaluate instream flows and connectivity using trail camera images associated with spatial location and time. MAP has demonstrated the effectiveness of this approach by evaluating flow regimes and connectivity with anthropogenic-based water withdrawals: https://doi.org/10.1002/rra.3689. Labeling thousands of images is time intensive however and even the best user interface will cause significant user eye fatigue.
We have previously explored a deep learning approach with minimal feature engineering to predict the appropriate label and are now pursuing development of multiple semi supervised weak-classifiers trained on features such as river bank lines, foliage bounding boxes and stream horizon lines to feed an ensemble classifier. We have implemented our prior methods using the python programming language relying upon the established open source OpenCV, piexif, tensorflow and scikit-learn packages for development: https://github.com/timothyjamesbecker/eco_image. Currently we have CNN models that perform well but will build new Vision Transformer models to compare using the new keras 3 API.
We plan to finish our preliminary work this Summer 2024 with the help of two paid Conn undergraduate student researchers that have been working on OpenCV projects for this past year in preparation. The result will be a manuscript as well as an open source github pip package for processing this type of river data and to use in ference for assistance with classifying new images.