Remote Sensing (Apr 2024)

Mapping Integrated Crop–Livestock Systems Using Fused Sentinel-2 and PlanetScope Time Series and Deep Learning

  • João P. S. Werner,
  • Mariana Belgiu,
  • Inacio T. Bueno,
  • Aliny A. Dos Reis,
  • Ana P. S. G. D. Toro,
  • João F. G. Antunes,
  • Alfred Stein,
  • Rubens A. C. Lamparelli,
  • Paulo S. G. Magalhães,
  • Alexandre C. Coutinho,
  • Júlio C. D. M. Esquerdo,
  • Gleyce K. D. A. Figueiredo

DOI
https://doi.org/10.3390/rs16081421
Journal volume & issue
Vol. 16, no. 8
p. 1421

Abstract

Read online

Integrated crop–livestock systems (ICLS) are among the main viable strategies for sustainable agricultural production. Mapping these systems is crucial for monitoring land use changes in Brazil, playing a significant role in promoting sustainable agricultural production. Due to the highly dynamic nature of ICLS management, mapping them is a challenging task. The main objective of this research was to develop a method for mapping ICLS using deep learning algorithms applied on Satellite Image Time Series (SITS) data cubes, which consist of Sentinel-2 (S2) and PlanetScope (PS) satellite images, as well as data fused (DF) from both sensors. This study focused on two Brazilian states with varying landscapes and field sizes. Targeting ICLS, field data were combined with S2 and PS data to build land use and land cover classification models for three sequential agricultural years (2018/2019, 2019/2020, and 2020/2021). We tested three experimental settings to assess the classification performance using S2, PS, and DF data cubes. The test classification algorithms included Random Forest (RF), Temporal Convolutional Neural Network (TempCNN), Residual Network (ResNet), and a Lightweight Temporal Attention Encoder (L-TAE), with the latter incorporating an attention-based model, fusing S2 and PS within the temporal encoders. Experimental results did not show statistically significant differences between the three data sources for both study areas. Nevertheless, the TempCNN outperformed the other classifiers with an overall accuracy above 90% and an F1-Score of 86.6% for the ICLS class. By selecting the best models, we generated annual ICLS maps, including their surrounding landscapes. This study demonstrated the potential of deep learning algorithms and SITS to successfully map dynamic agricultural systems.

Keywords