Using satellite imagery and mask R-CNN to monitor the Florida manatee (Trichechus manatus latirostris) population

Accurately monitoring the populations of endangered species is crucial to their conservation, and with limited time and funding available, it is imperative to find the most efficient way to do so. Aerial synoptic surveys, the current method for monitoring the population of the endangered Florida manatee, are costly, time-consuming, and tend to grossly underestimate the population. We propose training a Mask Region-Based Convolutional Neural Network (Mask R-CNN) to detect manatees from satellite imagery as a faster, less expensive, and more accurate alternative for monitoring the Florida manatee population. We build a Mask R-CNN using the TensorFlow Object Detection Application Programming Interface and pretrained weights from the Microsoft Common Objects in Context dataset. From Google Earth Pro, we generate 40 satellite images of manatees from along the same transect lines as the aerial surveys and annotate the ground truth in each image. We use 34 images to train the model and six images to test the model. The model performed with a mean average precision of 0.33. To improve the precision moving forward, we will collect more images of a higher resolution and will increase the distance above sea level to include more of the surrounding environment. With our model we will determine a one-year population estimate for the Florida manatee, which we will compare to the population estimate from that year’s aerial survey. We will communicate our results to the Florida Fish and Wildlife Service in hopes that they will incorporate our findings as they determine how best to survey the Florida manatee population in the future.

Hi, my name is Margaretta Walker. Under the guidance of my mentor, PhD candidate Sarah Farinelli, this summer I studied the possibility of using satellite imagery and a Mask Region-Based Convolutional Neural Network, or Mask R-CNN, to monitor the Florida manatee population. Florida manatees are an endangered species, and it is critical to monitor the populations of endangered species to assess threats, determine which management actions should be implemented, and then to test the effectiveness of those actions. Time and funding are precious resources in the field of conservation, so determining the most efficient way to survey threatened species is critical. Since 1991, the method used to survey the Florida manatee population has been aerial surveys, which involve human observers counting manatees from an aircraft at roughly 300 meters above sea level. Aerial surveys are costly and time-consuming, however, and up until 2014 Florida’s manatee surveys did not account for missed manatees. Individuals can be missed due to some manatees migrating out of Florida in warmer months, the large habitat area to cover in Florida, and manatees often being too far below the water surface to be seen from aircraft. This can result in gross underestimations of the manatee population, making it difficult to accurately assess threats and determine effective management actions. In recent years, satellite imagery has emerged as a faster, cheaper population survey method that is easier in remote areas; however, it has not yet been used to survey the Florida manatee. We propose the use of satellite imagery, with automated manatee detection by a Mask R-CNN, as an alternative to aerial surveys for monitoring the Florida manatee population. Convolutional Neural Networks, or CNNs, are a type of machine learning algorithm which can learn characteristics of a particular object from a set of images and then use those characteristics to make predictions on new images. Automated detection of animals from satellite imagery by a CNN is cheaper and faster than manual detection by humans. It is also more efficient for long-term monitoring and in hard-to-access locations. Additionally, once the CNN is taught to identify a species in one location, this knowledge can be transferred to other locations. We use a type of CNN called a Mask R-CNN, which has been shown to work well with over-lapping and close-together objects, such as aggregations of manatees, as well as with objects that are similar in color or texture to the background, as manatees are to water. Mask R-CNN divides each image into regions, and within each region creates a bounding box around the object and then a “mask” within that box which covers the object. Aerial surveys are conducted in Florida during January and February when manatees gather at warm water springs. Using the satellite imagery available in Google Earth Pro, I searched along the same transect lines as the 2018 and 2019 aerial surveys at roughly 150 meters above sea level, locating manatees and drawing polygons around them. From these polygons, I generated 40 images of manatees from a distance above sea level close enough that the shape of the manatees’ bodies could be distinguished but not so close that the images became distorted. I resized the images so that all 40 had the same dimensions and so that no one manatee appeared in more than one image. I then used the VGG Image Annotator software to manually label the ground truth in each image, or to create regions around the outline of each manatee. We divided the labelled images into a set of 34 images to train the model and a set of six images to test the model. Another researcher on the project wrote the code to build the model using the TensorFlow Object Detection Application Programming Interface and pretrained weights, or predicted probabilities that an object occurs in a given region, from the Microsoft Common Objects in Context dataset. After being trained, the model was tested using the validation image set. The third image here is an example of the resulting prediction. As you can see, the model correctly detected one of the two manatees. The 0.915 is the confidence score, which means the model is 91.5% sure that a manatee occurs in that particular bounding box. Overall, the model performed with a mean average precision of 0.33, which means that 33% of the time the model correctly guessed the location of a manatee as compared to the ground truth. This sounds low, however, given the quality of our satellite images and the limited amount of the surrounding environment included in each image, the model performs well. Moving forward, we will collect more images of a higher resolution and will increase the distance above sea level to include more of the surrounding environment. With these changes, we expect a model with significantly greater performance and precision. From there we will test our model on the Florida Application Programming Interface to determine a one-year population estimate for the Florida manatee, which we will compare to the population estimate from that year’s aerial survey. We will communicate our results to the Florida Fish and Wildlife Service in hopes that they will incorporate our findings as they determine how best to survey the Florida manatee population in the future. Thank you.