Search

The use of neural network (NN) models for remote sensing (RS) retrieval of landscape biophysical and biochemical properties has become popular in the last decade. Recently, the emergence of "big data" that can be generated from remotely sensed data and innovative machine learning (ML) approaches have provided a platform for novel analytical approaches. Specifically, the advent of deep learning (DL) frameworks developed from traditional neural networks (TNN) offer unprecedented opportunities to improve the accuracy of SOC retrievals from remotely sensed imagery. This review highlights the use of TNN models and their evolution into DL architectures in remote sensing of SOC estimation. The review also highlights the application of DL, with a specific focus on its development and adoption in remote sensing of SOC mapping. The review concludes by highlighting future opportunities for the use of DL frameworks for the retrieval of SOC from remotely sensed data. [ABSTRACT FROM AUTHOR]

Parthenium weed (Parthenium hysterophorus) is one of the most noxious herbaceous weeds in the world with adverse impacts on among others animal and human health, crop production, the environment, local as well as national economies. To optimize Parthenium mitigation, it is necessary to accurately monitor its spread using earth observation data. However, one of the challenges of mapping Parthenium weed is that its spectral response is similar to that of surrounding herbaceous plant species, resulting in low classification accuracies. Due to variability in its phenological characteristics and associated species, determining differences within the growing season may optimize the discrimination and subsequent mapping of the Parthenium weed. However, determination of the window(s) with the most prominent variability has been overlooked in past studies. Furthermore, no specific algorithm has been determined to be efficient in finding such window(s). ExtraTrees (EXT), an underused classifier in earth observation studies, possess interesting properties for satellite image processing such as high speed and performance. In this regard, this study attempted to (1) determine the optimal window period for discriminating Parthenium weed from coexisting plant species and (2) to compare the performance of EXT and the random forest (RF) algorithms. Results showed that the beginning of February was the optimal period for mapping Parthenium weed, with overall accuracy of 88.1%. EXT outperformed RF for most of the dates. This study lays the foundation for optimizing earth observation data derived models for characterizing invasive species, leveraging on the high temporal resolution of the new generation sensors. [ABSTRACT FROM AUTHOR]