A generalizable normalization for assessing plant functional diversity metrics across scales from remote sensing

Abstract

Remote sensing (RS) increasingly seeks to produce global‐coverage maps of plant functional diversity (PFD) across scales. PFD can be quantified with metrics assessing field or RS data dissimilarity. However, their comparison suffers from the lack of normalization approaches that (1) correct for differences in the number and correlation of traits and spectral variables and (2) do not require comparing all the available samples to estimate the maximum trait's dissimilarity (unfeasible in RS). We propose a generalizable normalization (GN) based on the maximum potential dissimilarity for the traits and spectral data considered and compare it to more traditional approaches (e.g. the maximum dissimilarity within datasets). To do so, we simulated plant communities with radiative transfer models and compared RS‐based diversity measurements across spatial scales (α‐ and β‐diversity components). Specifically, we assessed the capability of different normalization approaches (GN, local, none) to provide PFD estimates comparable between (1) RS and plant traits and (2) estimates from different RS missions. Unlike the other approaches, GN provides diversity component estimates that are directly comparable between field data and RS missions with different spectral configurations by removing the effect of differences in the number of traits or bands and the maximum dissimilarity across datasets. Therefore, GN enables the separated analysis of RS images from different sensors to produce comparable global‐coverage cartography. We suggest GN is necessary to validate RS approaches and develop interpretable maps of PFD using different RS missions.

Empowering Agribusiness with Simplicity: Web-based Artificial Intelligence Software Platform for Remote Sensing

“…It is believed that this is due to the fact that classifiers perform well with a large amount of data, and when indices are applied, the amount of data is reduced to a few values, suppressing many details of the samples in theory. It is believed that the relatively small number of images (11,200) used for training the neural network algorithms had a negative impact on the accuracy of this classifier. It is believed that a larger quantity of images per class, as long as they are balanced, can enable neural networks to outperform other classifiers that are not based on artificial neurons, potentially achieving accuracy rates above 90%.…”