Remote sensing of crop biophysical and biochemical parameters using sentinel-2 and machine learning algorithms

Abstract

Globally, achieving food security is critical to eliminating hunger, poverty, and malnutrition and reducing food-related diseases and deaths. This is enshrined in the United Nations’ Sustainable Development Goals (SDGs), African Union’s Agenda 2063, and national policies. The SDG 2 Target 2.4 aims to increase agricultural productivity and resilience to climate-related disasters, while also maintaining the integrity of agroecosystems. It is recommended to adopt innovative technologies and sustainable agricultural management practices to avert many challenges facing agricultural productivity and sustainability in food-insecure regions such as sub-Saharan Africa. This study sought to characterise the field variability in relevant crop biophysical and biochemical parameters for precision agriculture using the new generation sensor, Sentinel-2, and machine learning algorithms. Moreover, pressing issues regarding residual errors after atmospheric correction and their impact on the retrieval of crop parameters were addressed. The critical findings from this study are: (1) atmospheric correction (AC) techniques perform poorly under partly cloudy conditions, especially in the visible and SWIR region of the electromagnetic spectrum; (2) while it is beneficial to perform AC before estimating leaf and canopy chlorophyll content, LAI can be obtained directly from Top-of-Atmosphere (TOA) reflectance; (3) the study ascertained that there is no universal retrieval algorithm; hence, many have to be tested for specific sites, phenology, crop types, and crop parameters; (4) Sentinel-2 10 m (VNIR) and 20 m (RE-SWIR) bands can be used independently—without the need for superresolving techniques or spatial resampling that may cause a ~7 min computational delay; (5) the incorporation of per-pixel view and illumination geometries to spectral bands and vegetation indices improves the accuracy of retrieving essential crop parameters; (6) the spatial transferability of the retrieval model can be improved by adding limited (new or unseen) samples from the target site to account for variability in that site; and (7) a novel Spectral Triad feature selection technique, developed in the current study, was capable of optimising Sentinel-2 Multi-spectral Imager (MSI) feature space (and performed better than the entire feature space and v had similar performance with a well-established algorithm, i.e., Recursive Feature Elimination). The findings contribute to the existing knowledge of remotely sensed crop parameters and have practical implications for precision agriculture.