PS 37-203 - Detection of aquatic invasive species in wetlands using deep learning technologies and very high-resolution UAV imagery

Background/Question/Methods

Aquatic Invasive Plants (AIPs) are a global threat to local biodiversity due to their rapid adaptation to the new environments. They benefit from ecosystem changes and habitat disturbances by climatic changes and anthropogenic impacts. Lythrum salicaria, commonly known as Purple Loosestrife, is a predominant AIP that has invaded all counties in Wisconsin and has been designated as a deadly threat to the state’s wetlands. Accurately estimating the current extent of AIPs and slowing their future spread has been identified as a top priority for the WI-DNR. However, regular monitoring and detection is limited due to the use of expensive labor- and time-intensive field surveys. The very high-resolution imagery from Unmanned aerial vehicles (UAV) and computer vision technologies like Deep Learning (DL), also known as self-learning artificial intelligence approach, has shown promising results in automatic detection of different invasive species in wetland ecoregions and allow integration of expert knowledge and machine learning to accomplish the task. The objectives of this research are: (i) to develop a DL model for semantic segmentation and classification to identify Purple Loosestrife patches, and (ii) create an open-source machine learning pipeline for repeat implementation by the natural resource managers.

Results/Conclusions

A U-net based deep CNN model was implemented for semantic segmentation of Purple Loosestrife patches on Google Cloud Compute platform using open-source python packages with Keras API and TensorFlow backend. The model used UAV imagery (RGB, 2 cm pixel resolution) of the La Crosse River delta near Bangor, Wisconsin. The data was collected during the peak flowering season (July-August 2019) when Purple Loosestrife is most easily distinguishable from other native aquatic plants. The training and testing data were developed by visual interpretation of images. Among the different modeling architectures tested, ResNet152 provided the best result. The model was trained in 100 epochs with batch size 12. Jaccard loss function was used to calculate loss and Intersection Over Union (IOU) was used as accuracy metric. The IOU score for the model was 0.91. The predicted model output was mosaiced to create a binary classified raster data which was compared with the visually delineated reference data which showed that the model under predicted and the error prone areas include shorter bushes within the shadow of taller ones and dried grass patches surrounding Purple Loosestrife bushes. In future, this pre-trained model will be used in another study area to test transfer learning approach.