ResNeTS: A ResNet for Time Series Analysis of Sentinel-2 Data Applied to Grassland Plant-Biodiversity Prediction

Abstract

Analyzing time series from remote sensing data can aid in understanding spectral-temporal phenomena in ecosystems, such as the seasonal variation of plant components. Lately, deep learning has emerged as a strong method for mapping environmental variables from this data due to its exceptional predictive capabilities. This work studies the adaptation of the ResNet computer vision architecture for time series analysis of Sentinel-2 data. The resulting deep learning architecture, ResNeTS, stacks sequential convolutions to build a deep and narrow network, aligning with the design principles of leading convolutional architectures in computer vision. Experiments were carried out for predicting different plant-biodiversity indices, namely, species richness, and Shannon and Simpson indices, for temperate grassland ecosystems. The results show that ResNeTS can achieve moderate improvements in terms of accuracy compared to other state-of-the-art architectures, such as InceptionTime (up to +0.021 $r^{2}$), with reduced computational costs owing to its streamlined architecture.

Keywords

• Biodiversity prediction
• deep learning
• multispectral imaging
• remote sensing
• residual network (ResNet)
• sentinel-2