Enhancing Ecological Monitoring with Multi-Objective Optimization: A Novel Dataset and Methodology for Segmentation Algorithms

TL;DR

The paper introduces a 6,096-image aerial dataset of native and invasive grasses from the Bega Valley and a homotopy-based multi-objective fine-tuning framework that balances segmentation accuracy with contextual coherence, demonstrated on SAM. By integrating DiceCELoss and a smoothness loss with a gradually evolving homotopy parameter, it achieves robust segmentation in noisy ecological imagery and provides strong baselines across multiple architectures. The dataset includes field-collected imagery, auxiliary metadata, and a simplified grass/non-grass annotation protocol, enabling large-scale ecological monitoring research. Results show the proposed Multi-Objective SAM achieves top performance with efficient training, highlighting its practical potential for early detection and management of invasive grasses like African lovegrass. The work advances ecological CV by offering a high-quality dataset, a novel training paradigm, and actionable insights for real-world environmental monitoring and sustainable land management.

Abstract

We introduce a unique semantic segmentation dataset of 6,096 high-resolution aerial images capturing indigenous and invasive grass species in Bega Valley, New South Wales, Australia, designed to address the underrepresented domain of ecological data in the computer vision community. This dataset presents a challenging task due to the overlap and distribution of grass species, which is critical for advancing models in ecological and agronomical applications. Our study features a homotopy-based multi-objective fine-tuning approach that balances segmentation accuracy and contextual consistency, applicable to various models. By integrating DiceCELoss for pixel-wise classification and a smoothness loss for spatial coherence, this method evolves during training to enhance robustness against noisy data. Performance baselines are established through a case study on the Segment Anything Model (SAM), demonstrating its effectiveness. Our annotation methodology, emphasizing pen size, zoom control, and memory management, ensures high-quality dataset creation. The dataset and code will be made publicly available, aiming to drive research in computer vision, machine learning, and ecological studies, advancing environmental monitoring and sustainable development.

Figures (4)

• Figure 1: Estimated distribution of African lovegrass (ALG) infestation in Australia, based on data from Queensland Primary Industries and Fisheries (2009).
• Figure 2: A 25x25m$^2$ plot established for ecological monitoring, including the study of native grasses and African lovegrass interactions.
• Figure 3: Each ROC curve represents a model's ability to distinguish between grass and non-grass areas, with the Area Under the Curve (AUC) and Equal Error Rate (EER) values indicated in the legend. The EER points are marked with red circles on each curve, showing the threshold where the false positive rate equals the false negative rate. The dashed diagonal line represents the performance of a random classifier (AUC = 0.50). The higher AUC and lower EER values indicate better performance.
• Figure 4: Comparison of segmentation results from all models on the same aerial image. The comparison highlights the varying levels of detail, precision, and generalization ability of each model in distinguishing grass from non-grass areas. Note that some models like Multi-Objective SAM tend to misclassify large portions of the image as grass, while others like DeepLabV3 ResNet 101 perform better at capturing fine details.