Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Grazing Detection using Deep Learning and Sentinel-2 Time Series Data

Aleksis Pirinen, Delia Fano Yela, Smita Chakraborty, Erik Källman

TL;DR

This paper addresses scalable detection of grazing activity at field boundaries using Sentinel-2 L2A time series. It develops a CNN–LSTM ensemble that processes masked field-polygon time-series data, with temporal dropout and majority voting across the last four time steps to improve robustness. Across five cross-validation splits, the approach achieves an average F1 of about 0.77 and high grazing recall (~0.90), enabling practical gains in inspection efficiency by prioritizing fields predicted as non-grazed (up to 17.2× more confirmed non-grazed sites under limited visit budgets). The work demonstrates that freely available coarse-resolution satellite data can effectively steer conservation inspections and land-use enforcement, with potential for extension through larger, more diverse datasets and pretraining strategies.

Abstract

Grazing shapes both agricultural production and biodiversity, yet scalable monitoring of where grazing occurs remains limited. We study seasonal grazing detection from Sentinel-2 L2A time series: for each polygon-defined field boundary, April-October imagery is used for binary prediction (grazed / not grazed). We train an ensemble of CNN-LSTM models on multi-temporal reflectance features, and achieve an average F1 score of 77 percent across five validation splits, with 90 percent recall on grazed pastures. Operationally, if inspectors can visit at most 4 percent of sites annually, prioritising fields predicted by our model as non-grazed yields 17.2 times more confirmed non-grazing sites than random inspection. These results indicate that coarse-resolution, freely available satellite data can reliably steer inspection resources for conservation-aligned land-use compliance. Code and models have been made publicly available.

Grazing Detection using Deep Learning and Sentinel-2 Time Series Data

TL;DR

This paper addresses scalable detection of grazing activity at field boundaries using Sentinel-2 L2A time series. It develops a CNN–LSTM ensemble that processes masked field-polygon time-series data, with temporal dropout and majority voting across the last four time steps to improve robustness. Across five cross-validation splits, the approach achieves an average F1 of about 0.77 and high grazing recall (~0.90), enabling practical gains in inspection efficiency by prioritizing fields predicted as non-grazed (up to 17.2× more confirmed non-grazed sites under limited visit budgets). The work demonstrates that freely available coarse-resolution satellite data can effectively steer conservation inspections and land-use enforcement, with potential for extension through larger, more diverse datasets and pretraining strategies.

Abstract

Grazing shapes both agricultural production and biodiversity, yet scalable monitoring of where grazing occurs remains limited. We study seasonal grazing detection from Sentinel-2 L2A time series: for each polygon-defined field boundary, April-October imagery is used for binary prediction (grazed / not grazed). We train an ensemble of CNN-LSTM models on multi-temporal reflectance features, and achieve an average F1 score of 77 percent across five validation splits, with 90 percent recall on grazed pastures. Operationally, if inspectors can visit at most 4 percent of sites annually, prioritising fields predicted by our model as non-grazed yields 17.2 times more confirmed non-grazing sites than random inspection. These results indicate that coarse-resolution, freely available satellite data can reliably steer inspection resources for conservation-aligned land-use compliance. Code and models have been made publicly available.
Paper Structure (9 sections, 3 figures, 2 tables)

This paper contains 9 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Dataset
  3. Data preprocessing
  4. Machine learning-ready dataset
  5. Machine learning approach
  6. Model training and inference
  7. Experimental results
  8. In practice: Application on grazing inspections
  9. Conclusions

Figures (3)

  • Figure 1: Example RGB-parts of Sentinel-2 L2A time series and field boundaries (polygons).
  • Figure 2: Training and validation polygons across Sweden.
  • Figure 3: Expected percentage of non-grazing sites found under different visitation percentages $p$, when selecting sites at random either (i) from all sites (orange), or (ii) from the sites predicted as non-grazed by our model (blue). We here assume that roughly $5\%$ of all sites are non-grazed. If the visitation percentage $p \leq 4\%$, then our ML-based approach (ii) finds 17.2x more non-grazing sites, on average.