Remote and Proximal Assessment of Plant Traits

The inference of functional vegetation traits from remotely sensed signals is key to providing efficient information for multiple plant-based applications and to solve related problems. Functional traits can be of morphological, biochemical, physiological, structural or phenological nature and reveal performance and vitality of plants. These traits, variables or physiological characteristics include, for instance, plant height [1], leaf area index [2], biomass [3], nutrient contents [4], water status [5], pigments [6], photosynthetic activity [7], disease resistance [8], yield prediction [9] or stress detection, being used by a diversity of ecological and agricultural applications. Spatial–temporal information of these traits is needed in numerous ecosystem applications, such as agriculture (crop breeding/monitoring/yield estimation) or biodiversity and landscape ecology. Therefore, the development and adaptation of fast, efficient, accurate and generic retrieval algorithms for quantification of vegetation traits, directly or indirectly from remote sensing observations, is required. In this context, variable data-driven as well as radiometric-data driven methods are continuously expanding over the time with rising remotely sensing data processing applications and studies. Drivers of this methodological expansion are, among others, the increase in computational power, the easier availability and increasing dimensionality of Earth observation (EO) data, and finally the ongoing progress in designing and planning spaceborne imaging spectroscopy missions. All methods have their respective limitations and advantages. Most importantly, they should be provided on appropriate platforms and (free-to-use) software tools, in order to be evaluated by plant physiologists, agronomists, remote sensing scientists and ecologists. With the ambition to collect current state-of-the-art research, the Special Issue (SI) of “Remote and Proximal Assessment of Plant Traits” aimed to address contributions about estimation of morpho-physiological and biochemical plant traits from EO data in agricultural and ecological contexts, supporting food security and sustainability. Overall, with 14 published studies including 12 research studies, one technical note and one review study, a wide range of applications was covered meeting the expectations of the call for papers. The majority of studies aimed at estimating functional vegetation traits at local and a few at regional scale within different ecosystems (mainly agriculture). Remote sensing of plants (crops) was performed at leafor canopy-levels by means of diverse sensors on field-level to spaceborne platforms, covering multiand hyperspectral resolutions and applying a multitude of retrieval algorithms. In addition, these quantification studies, Castrignanò et al. [8] explored a classification method for the detection of biotic stress and two studies used time series analysis for assessing phenology [10] and yield [1]. Table 1 provides an overview of all the studies including the respective meta-information.

The inference of functional vegetation traits from remotely sensed signals is key to providing efficient information for multiple plant-based applications and to solve related problems. Functional traits can be of morphological, biochemical, physiological, structural or phenological nature and reveal performance and vitality of plants. These traits, variables or physiological characteristics include, for instance, plant height [1], leaf area index [2], biomass [3], nutrient contents [4], water status [5], pigments [6], photosynthetic activity [7], disease resistance [8], yield prediction [9] or stress detection, being used by a diversity of ecological and agricultural applications. Spatial-temporal information of these traits is needed in numerous ecosystem applications, such as agriculture (crop breeding/monitoring/yield estimation) or biodiversity and landscape ecology. Therefore, the development and adaptation of fast, efficient, accurate and generic retrieval algorithms for quantification of vegetation traits, directly or indirectly from remote sensing observations, is required. In this context, variable data-driven as well as radiometric-data driven methods are continuously expanding over the time with rising remotely sensing data processing applications and studies. Drivers of this methodological expansion are, among others, the increase in computational power, the easier availability and increasing dimensionality of Earth observation (EO) data, and finally the ongoing progress in designing and planning spaceborne imaging spectroscopy missions. All methods have their respective limitations and advantages. Most importantly, they should be provided on appropriate platforms and (free-to-use) software tools, in order to be evaluated by plant physiologists, agronomists, remote sensing scientists and ecologists.
With the ambition to collect current state-of-the-art research, the Special Issue (SI) of "Remote and Proximal Assessment of Plant Traits" aimed to address contributions about estimation of morpho-physiological and biochemical plant traits from EO data in agricultural and ecological contexts, supporting food security and sustainability. Overall, with 14 published studies including 12 research studies, one technical note and one review study, a wide range of applications was covered meeting the expectations of the call for papers.
The majority of studies aimed at estimating functional vegetation traits at local and a few at regional scale within different ecosystems (mainly agriculture). Remote sensing of plants (crops) was performed at leaf-or canopy-levels by means of diverse sensors on field-level to spaceborne platforms, covering multi-and hyperspectral resolutions and applying a multitude of retrieval algorithms. In addition, these quantification studies, Castrignanò et al. [8] explored a classification method for the detection of biotic stress and two studies used time series analysis for assessing phenology [10] and yield [1]. Table 1 provides an overview of all the studies including the respective meta-information. The commonly accepted taxonomy of retrieval methods (for short overviews and appropriate references see Berger et al. [12], Estevez et al. [14] or Mahajan et al. [4]) was nicely reflected by the studies of our SI: (i) Parametric regressions, mainly referring to the use of different (hyperspectral) vegetation indices (VI) [3,4,6,7,9]; (ii) Nonparametric regressions, which include chemometric methods, such as partial least square regressions (PLSR), as well as machine learning (ML) regression algorithms, such as artificial neural networks, support vector regression (SVR) or Gaussian process regression (GPR) [4,6]; (iii) Physically based or inversion of leaf and canopy radiative transfer models (RTM), for instance the 1D approaches of PROSAIL [6] or 3D DART model families [5]; (iv) Hybrid approaches, where simulated RTM data serve for training of ML regression algorithms [2,12,14].
With the methodology covered by the published studies, we can recognize a clear trend towards more complex retrieval methods (ii-iv) moving away from the traditional and still most frequently used parametric regressions (i), towards a deeper understanding of underlying physical processes of the radiation-vegetation interactions by exploring RTMs (iii). Hybrid approaches (iv), combining RTMs with efficient and fast ML regression algorithms, emerged as most promising category as demonstrated by three studies [2,12,14]. These methods could become the key player of next-generation retrieval strategies in respect of a potential routine delivery of global vegetation products by future spaceborne imaging spectroscopy missions. For instance, top-atmosphere retrieval from Sentinel-2 data of multiple vegetation traits was suggested by Estévez et al. [14], which could be extended to spaceborne hyperspectral data in the future. Moreover, the implementation of intelligent sampling methods (active learning) to provide faster and lighter retrieval models, particularly useful for implementation into cloud computing platforms, such as the Google Earth Engine (GEE), was proposed by Berger et al. [12]. In this context, an appealing solution for further reduction of computation times was presented by Miraglio et al. [5]: Hereby, reflectance curves from the complex 3D DART model were approximated using the simpler and faster 1D PROSAIL model, with this drastically reducing model running time for generation of a reflectance database. According to Mahajan et al. [4], the combination of different nonparametric approaches can also provide efficient, elegant and highly accurate solutions for the prediction of multiple plant nutrients from hyperspectral data, as the authors nicely demonstrated with the synergistic use of PLSR and SVR models. Moreover, the studies of Aharon et al. [11] used image analyses and 3D modeling to assess morphological traits for weed competitiveness and recommended for future work to add machine learning techniques. Wang et al. [10] analyzed imagery at larger spatial (county) and temporal (16 years) scales to assess phenology in relation to weather conditions based on a time series approach. To complete the scale range, a close range phenotyping system based on sensing techniques was developed by Khak Pour et al. [13], who presented their high throughput data collection platform with an exemplary wheat monitoring case study.
Regarding sensor systems, a decisive trend of increasing UAV data exploitation could be identified. This was particularly the case for the prediction of agronomic traits within breeding programs [1,9] and stress detection [8], but also for managing crop production [3]. Except for one study [12], simulating data from the German environmental mapping and analysis program (EnMAP), spaceborne imaging spectroscopy data were not explored by the published papers. This is about to change, with increasing data streams to be provided by current (PRISMA) or near-term missions, such as EnMAP, or future CHIME. Moreover, the synergistic usage of multiple optical sensors systems could greatly enhance the information content in respect to the different research questions as addressed in the published articles. Still, some other gaps and challenges remain, such as handling of large amounts of data, dealing with spatial and temporal structures in a more efficient way and also improvements of retrieval accuracy in particular of the leaf-level traits. To address this, an attractive option could be the exploitation of deep learning methods.
In summary, the 14 studies published here not only reflect the continuous advancement of new Earth observation sensors and methods, but also may support in further stimulating the ongoing progress in the field of vegetation traits retrieval and related applications. This ongoing progress will contribute towards more efficient operational monitoring of plants and crops, from plant organs, small plots, sub-field, field over regional to global scales, providing breeders, farmers and ecologists with relevant information for their decisionmaking processes in the context of Agriculture 4.0, as well as for additional research.