Evaluating Convolution and Attention Time Series Algorithms for Classification in Tropical Biomes

Le résumé fourni par la source

This study compares convolution-based and attention-based deep learning algorithms for land classification in tropical biomes using satellite image time series. We used two contrasting study areas to reduce spatial bias: Rondonia in the Amazon rainforest biome and Petrolina in the Caatinga semi-arid biome. We chose the Temporal Convolutional Neural Network (TCNN) and Lightweight Temporal Self-Attention (LTAE) algorithms. In the Amazon, the TCNN model achieved higher accuracies when detecting deforestation events. The LTAE model achieved the best accuracy in the Caatinga since its global attention structure is better suited to capture unstructured temporal signatures. We also compared pixel-based and region-based versions of these algorithms and found that pixel-based versions perform better. Our results show that the choice of DL algorithms for land classification is problem-dependent.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.