[Paper Review] Counting Cows: Tracking Illegal Cattle Ranching From High-Resolution Satellite Imagery

Issam Laradji, Pau Rodríguez+5 more|arXiv (Cornell University)|Nov 14, 2020

Video Surveillance and Tracking Methods32 references18 citations

TL;DR

This paper proposes a deep learning approach to detect and count cattle in high-resolution satellite imagery (0.31 m) across the Amazon, using CSRNet and LCFCN models to enable large-scale monitoring of illegal cattle ranching. Results show CSRNet outperforms LCFCN in dense regions, achieving low MAPE and GAMPE, demonstrating the feasibility of automated, scalable tracking of illegal land use via commercial satellite data.

ABSTRACT

Cattle farming is responsible for 8.8\% of greenhouse gas emissions worldwide. In addition to the methane emitted due to their digestive process, the growing need for grazing areas is an important driver of deforestation. While some regulations are in place for preserving the Amazon against deforestation, these are being flouted in various ways, hence the need to scale and automate the monitoring of cattle ranching activities. Through a partnership with extit{Global Witness}, we explore the feasibility of tracking and counting cattle at the continental scale from satellite imagery. With a license from Maxar Technologies, we obtained satellite imagery of the Amazon at 40cm resolution, and compiled a dataset of 903 images containing a total of 28498 cattle. Our experiments show promising results and highlight important directions for the next steps on both counting algorithms and the data collection process for solving such challenges. The code is available at \url{https://github.com/IssamLaradji/cownter_strike}.

Motivation & Objective

• To develop an automated, scalable method for detecting and counting cattle in high-resolution satellite imagery to monitor illegal ranching in the Amazon.
• To evaluate the performance of deep learning models—specifically CSRNet and LCFCN—for cattle counting and localization at continental scale.
• To assess the feasibility of using commercial sub-meter satellite imagery (0.31 m resolution) for large-scale environmental monitoring.
• To identify limitations in current approaches and propose improvements through multi-temporal imagery and better annotation strategies.

Proposed method

• Utilized 903 high-resolution satellite images (0.31 m panchromatic, 1.24 m multi-spectral) from Maxar’s WorldView3, acquired via task-based scheduling.
• Compiled a dataset of 28,498 cattle with point-level annotations for training and evaluation.
• Trained CSRNet (density-based) and LCFCN (detection-based) models using VGG16-FCN8 backbones pre-trained on ImageNet.
• Employed a detection-based loss function in LCFCN that encourages single-blob outputs per object, improving localization and robustness to scale variation.
• Used MAPE and GAMPE as evaluation metrics, with GAMPE partitioning images into grids to penalize spurious correlations.
• Applied early stopping and ADAM optimization with a fixed initial learning rate of 10⁻⁴ over 100 epochs.

Experimental results

Research questions

• RQ1 Can high-resolution satellite imagery (0.31 m) enable accurate detection and counting of cattle at continental scale?
• RQ2 How do density-based (CSRNet) and detection-based (LCFCN) deep learning models compare in cattle counting performance across varying cattle densities?
• RQ3 To what extent does pre-training on ImageNet improve model generalization for cattle counting in remote sensing applications?
• RQ4 Can multi-temporal satellite revisits improve the distinction between cattle and similar-looking static objects (e.g., rocks, bushes)?

Key findings

• CSRNet achieved superior performance over LCFCN in both counting and localization, especially in dense regions (>100 cattle per image), with lower MAPE and GAMPE.
• LCFCN outperformed CSRNet in the absence of cattle (0 per image), achieving an F-score of 0.676 compared to CSRNet’s 0.571.
• The ImageNet-pretrained LCFCN outperformed the non-pretrained version, indicating benefit from transfer learning.
• The hybrid use of CSRNet and LCFCN is promising, with CSRNet excelling in dense counting and LCFCN in binary presence/absence detection.
• Multi-temporal imagery is critical for improving labeling quality and distinguishing cattle from false positives such as rocks and bushes.
• The study confirms that commercial 0.31 m satellite sensors are capable of reliable cattle counting, enabling scalable environmental monitoring.