Coastal marshlands are vital ecosystems, providing essential services such as storm protection and habitat for diverse wildlife. However, monitoring their health is a logistical challenge. Organizations like the Guana Tolomato Matanzas National Estuarine Research Reserve (GTMNERR) utilize "photoquadrats"—high-resolution, one-meter-square images of the ground—to estimate "percent cover." Currently, this process requires volunteers to manually identify the contents of 100 randomly selected pixels per image, creating a massive bottleneck in environmental monitoring. This research aims to automate this process using deep learning.
Methodology
The researchers developed a multi-stage pipeline designed to mimic the manual identification workflow. The process begins by selecting 100 random points from a full-size marsh photograph. To provide necessary context for classification, the system extracts a $31 \times 31$ pixel "context window" around each point rather than analyzing individual pixels in isolation.
These snippets are fed into a Convolutional Neural Network (CNN) to determine if the center pixel is "Vegetated" or "Unvegetated." The study compared two classic architectures: LeNet-5 and a modified AlexNet. To optimize for speed and the specific constraints of the dataset, the authors reduced the number of units in AlexNet’s fully connected layers to 100. The models were trained using a stratified 80/20 train-test split with 10-fold cross-validation.
Dataset Development
Because existing records lacked specific pixel coordinates for previous labels, a new ground-truth dataset was built from scratch. Using photoquadrats from St. Augustine, Florida, a custom labeling program allowed volunteers to tag 57,410 snippets. Each point was categorized as either unvegetated or one of five specific vegetation species.
Key Results and Insights
The binary classification task (Vegetated vs. Unvegetated) yielded strong results, with AlexNet emerging as the preferred architecture:
- AlexNet: Achieved a test accuracy of 92.41% (training accuracy: 99.48%).
- LeNet-5: Achieved a test accuracy of 91.34% (training accuracy: 99.05%).
Analysis of the learning curves suggests that while both models performed well, they showed signs of underfitting. The authors noted a discrepancy between validation and test accuracy, attributed to "underrepresented classes"—certain plant species or soil types that appear less frequently in the training data. However, the results confirm that context is essential; the $31 \times 31$ snippets allow the models to recognize the textures and shapes of grass blades and soil that are critical for accurate classification in messy natural environments.
Future Work
The current system acts as a binary gatekeeper, but several expansions are planned:
- Species Identification: Expanding the model into a 6-class classifier to distinguish between specific marsh grass species.
- Dataset Diversity: Increasing sample sizes for underrepresented classes to improve model confidence.
- Ensemble Modeling: Using multiple models to "vote" on classifications to increase reliability.
- Deployment: Developing a Django-based web application to allow researchers to upload images and receive automated reports in seconds, transforming a hours-long manual task into an automated one.
Summary updated