Our understanding of the Earth´s functional biodiversity and its imprint on ecosystem functioning is still incomplete. Large-scale information on functional ecosystem properties (‘Plant Traits’) is thus urgently needed to assess functional diversity and better understand biosphere-environment interactions. Optical remote sensing and particularly hyperspectral data offer a powerful tool to map these biophysical properties. Such data enable repeatable and non-destructive measurements at different spatial and temporal scales over continuous narrow bands and using numerous platforms and sensors. The advent of the upcoming space-borne imaging spectrometers will provide an enormous amount of data that opens the door to explore data driven methods for processing and analysis. However, we are still lacking until now efficient and accurate methods to translate hyperspectral reflectance into information on biophysical properties across plant types, environmental gradients and sensor types. In this regard, Deep Learning (DL) techniques are revolutionizing our capabilities to exploit large data sets given their flexibility and efficiency to detect features and their complex and hierarchical relationships. Accordingly, it is expected that Convolutional Neural Networks (CNNs) have the potential to provide transferable predictive models of biophysical properties at the canopy scale from spectroscopy data. On the other side, the absence of globally representative data sets and the gap between the available reflectance data and the corresponding in-situ measurements are reasons that hampered such analyses until now. In recent years, several initiatives from the scientific community (e.g. EcoiSIS) have contributed to provide a constantly growing source of data of hyperspectral reflectance and plant trait encompassing different plant types and sensors. However, such data are sparse to fit whatever model because of missing values. In the present study, we demonstrate a weakly supervised approach to enrich these data sets using gap filling strategies. Based on this data, we investigate different multi-output Deep Learning (DL) architectures in a form of an end-to-end workflow that predicts multiples biophysical properties at once. Based on 1D-CNN the model exploits the internal correlation between multiple traits and hence improves predictions. In the study, we target a various set of plant properties from pigments, structural traits (e.g. LAI), water content, nutrients (e.g. Nitrogen) and Leaf mass area (LMA). The preliminary results of the mapping model cross a broad range of vegetation types (Crops, Forest, Tundra, Grassland) are promising and outcompete the performance of shallow machine learning approaches (e.g. Partial Least Squares Regression (PLSR), Random Forest Regression) that can only predict individual traits. The model learned distinguishable and generalized features despite of the high variability in the used data sets. The key contribution of this study is to highlight the potential of weakly supervised approaches together with Deep Learning to overcome the scarcity of in-situ measurements and take a step forward in creating efficient predictive models of multiple Earth’s biophysical properties.