[Paper Review] Detecting Cattle and Elk in the Wild from Space

Caleb Robinson, Anthony Ortiz+3 more|arXiv (Cornell University)|Jun 29, 2021

Wildlife Ecology and Conservation20 references4 citations

TL;DR

This paper proposes CowNet, a deep learning model that simultaneously counts and localizes cattle and elk in very high-resolution satellite imagery using a shared spatial representation. It achieves state-of-the-art performance with LC-FCN model attaining 0.56–0.61 average precision and 0.78–0.92 average recall on held-out test scenes, while introducing novel evaluation methods accounting for label noise and large-scale scene analysis.

ABSTRACT

Localizing and counting large ungulates -- hoofed mammals like cows and elk -- in very high-resolution satellite imagery is an important task for supporting ecological studies. Prior work has shown that this is feasible with deep learning based methods and sub-meter multi-spectral satellite imagery. We extend this line of work by proposing a baseline method, CowNet, that simultaneously estimates the number of animals in an image (counts), as well as predicts their location at a pixel level (localizes). We also propose an methodology for evaluating such models on counting and localization tasks across large scenes that takes the uncertainty of noisy labels and the information needed by stakeholders in ecological monitoring tasks into account. Finally, we benchmark our baseline method with state of the art vision methods for counting objects in scenes. We specifically test the temporal generalization of the resulting models over a large landscape in Point Reyes Seashore, CA. We find that the LC-FCN model performs the best and achieves an average precision between 0.56 and 0.61 and an average recall between 0.78 and 0.92 over three held out test scenes.

Motivation & Objective

• To develop a deep learning-based method for simultaneous counting and localization of large ungulates in very high-resolution satellite imagery.
• To evaluate model performance on large scenes while accounting for spatial noise in human-labeled data.
• To benchmark temporal generalization of models across time series imagery in a real-world ecological monitoring context.
• To propose evaluation metrics tailored for ecological applications, emphasizing recall and robustness to label uncertainty.
• To identify the most effective model architecture for scalable wildlife monitoring using satellite data.

Proposed method

• CowNet uses a shared encoder-decoder architecture to jointly predict object counts and pixel-level localization maps from VHR satellite imagery.
• The model is trained using a multi-task loss combining count regression and localization supervision via point annotations.
• A novel evaluation framework computes precision and recall using a cutoff distance to account for label noise, with sensitivity analysis used to select optimal thresholds.
• Models are evaluated on a grid-based resolution to assess performance across varying spatial scales, with metrics computed at 100m, 256m, and 1024m cell sizes.
• Temporal generalization is tested by training on early scenes (2013–2014), validating on 2016, and testing on 2015–2017 scenes to simulate real-world deployment.
• The study compares multiple architectures including LC-FCN, U-Net, CSRNet, FCRN, and CowNet, using both density-based and segmentation-based prediction heads.

Experimental results

Research questions

• RQ1 How well can deep learning models simultaneously count and localize cattle and elk in very high-resolution satellite imagery?
• RQ2 How does model performance vary when evaluated on large, continuous scenes rather than fixed image patches?
• RQ3 What is the impact of label noise on performance estimation, and how can evaluation metrics be adapted to reflect this?
• RQ4 Which model architecture generalizes best across time when trained on a single region and tested on later imagery?
• RQ5 How does grid cell size affect the reliability of counting metrics in large-scale ecological monitoring?

Key findings

• The LC-FCN model achieves the highest performance, with average precision between 0.56 and 0.61 and average recall between 0.78 and 0.92 across three held-out test scenes.
• Performance improves with larger grid cell sizes due to cancellation of false positives and false negatives, with CowNet reaching R² = 0.82 at 1024m resolution.
• The CSRNet model only shows positive R² performance when grid cells exceed 256m, indicating poor localization at fine scales due to coarse density estimation.
• U-Net and LC-FCN show high variability between conservative and optimistic recall estimates, indicating large connected components in predictions that affect matching accuracy.
• A cutoff distance of d=4 meters was selected as optimal, as it captures most true positives affected by label noise while showing diminishing returns beyond d=5.
• The sensitivity analysis revealed a sharp performance increase between d=1 and d=3 meters, highlighting the significance of label misalignment in real-world data.