@Article{LucianoPiDuRoLeMa:2021:EmMoFo,
author = "Luciano, Ana Cl{\'a}udia dos Santos and Picoli, Michelle Cristina
Ara{\'u}jo and Duft, Daniel Garbellini and Rocha, Jansle Vieira
and Leal, Manoel Regis Lima Verde and le Maire, Guerric",
affiliation = "{Universidade de S{\~a}o Paulo (USP)} and {Instituto Nacional de
Pesquisas Espaciais (INPE)} and {Universidade de S{\~a}o Paulo
(USP)} and {Universidade Estadual de Campinas (UNICAMP)} and
{Universidade Estadual de Campinas (UNICAMP)} and Eco\&Sols, Univ
Montpellier, CIRAD, INRA, IRD",
title = "Empirical model for forecasting sugarcane yield on a local scale
in Brazil using Landsat imagery and random forest algorithm",
journal = "Computers and Electronics in Agriculture",
year = "2021",
volume = "184",
pages = "e106063",
month = "May",
keywords = "Crop yield, Remote sensing, Vegetation indices, Machine
learning.",
abstract = "Sugarcane plays an important role in food and energy production in
Brazil and worldwide. The large availability of satellite sensors
and advanced techniques for processing data have improved the
forecasting sugarcane yield on a local and global scale, but more
work is needed on exploiting the synergy between remote sensing,
meteorological and agronomic data. In this study, we combined such
data sources to forecast sugarcane yield using a random forest
(RF) algorithm on an extensive area of 50,000 ha, over four years.
Images from Landsat satellites were processed to time series of
surface reflectance and spectral indices. The approach focused on
the development of predictive models which only used data acquired
and accessible several months before the harvest. First, three RF
models were calibrated with different predictors to forecast the
sugarcane yield at harvest: using Landsat satellite images and
meteorological data (RF1); agronomic and meteorological data
(RF2); a combination of Landsat satellite images, agronomic and
meteorological data (RF3). As a comparison, we also tested the
influence of including knowledge on the future harvest date in the
models RF2 and RF3 (RF4 and RF5). The average values of R2 for
RF1, RF2, and RF3 were 0.66, 0.50 and 0.74, respectively. The
model with the highest values of R2 (RF3) had a Root Mean Square
Error (RMSE) of 9.9 ton ha\−1 on yield forecast,
approximately 15% of the yield average. Including the harvest date
improved the RF2 and RF3 models to reach R2 = 0.69 and RMSE = 10.8
ton ha\−1 for RF4, and R2 = 0.76 and RMSE of 9.4 ton
ha\−1 for RF5. A blind forecasting test for the 2016 yields
showed similar prediction than the forecast made by in situ field
expertise. This result has the potential to assist management of
sugarcane production.",
doi = "10.1016/j.compag.2021.106063",
url = "http://dx.doi.org/10.1016/j.compag.2021.106063",
issn = "0168-1699",
language = "en",
targetfile = "1-s2.0-S0168169921000818-main.pdf",
urlaccessdate = "20 maio 2024"
}