Leveraging Ground Sensor Networks to Evaluate Satellite-Based Land Surface Phenology in Smallholder Farming Systems

Abstract

Satellite-derived estimates of land surface phenology (LSP) are important for quantifying the spatiotemporal dynamics of agricultural landscapes where LSP metrics can track crop growth and productivity. Previous LSP applications for agriculture have mostly been confined to developed countries, where fields are larger and more homogenous, and ground data networks are available. LSP studies from developing regions are limited and have focused on coarse scales without field-level comparison. This study used field-level multispectral sensors (Arable Marks) to quantify several key LSP metrics and compare satellite and ground-derived LSP timing in smallholder maize fields. The ground sensors were installed in crop fields in Kenya and Zambia that were managed according to typical smallholder practices, and used to estimate greenup, maturity, senescence, and dormancy dates. We then calculated differences between the ground LSP metrics and corresponding metrics from Sentinel-1, Sentinel-2, and VIIRS time-series, and the VIIRS Land Cover Dynamics product. We also compared the satellite-based LSP metrics to farmer planting and harvest dates for a subset of fields. Results showed that Sentinel-2 had the smallest differences among single sensor models relative to ground-based LSP metrics across all dates (bias-adjusted MAD 11-13 days), with multi-sensor models showing comparable correlation (Kendall’s t) with ground LSP metrics. Sentinel-1 and the VIIRS Land Cover Dynamics product were less comparable to ground-based measures due to high variability and missing observations, respectively. Across most sensors and models, correlations with ground LSP metrics were higher for late season dates (senescence, dormancy) compared to early season dates (greenup, maturity). For management events, single-sensor models had higher correlation for harvest date (highest t = 0.75), than planting date (highest t = 0.30). Overall the results show that satellite-derived LSP can estimate ground-based measures of LSP within a two week timeframe for the majority of sites. We identify multi-sensor models, extended ground sensor networks, and high-resolution satellite imagery as priorities for continued research of smallholder LSP monitoring. These results are the first evaluation of satellite-derived LSP metrics for smallholder agriculture comparing multiple satellite sensors and vegetation indices and provide a baseline performance for global, freely available satellite sensors.

Keywords: smallholder, agriculture, land surface phenology, maize, ground sensor, multi- sensor, Zambia, Kenya