Maize Yield Estimation in Kenya Using Modis

Abstract

Abstract. Monitoring staple crop production can support agricultural research, business such as crop insurance, and government policy. Obtaining accurate estimates through field work is very expensive, and estimating it through remote sensing is promising. We estimated county-level maize yield for the 37 maize producing countries in Kenya from 2010 to 2017 using Moderate Resolution Imaging Spectroradiometer (MODIS) data. Support Vector Regression (SVR) and Random Forest (RF) were used to fit models with observed county level maize yield as a function of vegetation indices. The following five MODIS vegetation indices were used: green normalized difference vegetation index, normalized difference vegetation index, normalized difference moisture index, gross primary production, and fraction of photosynthetically active radiation. The models were evaluated with 5-fold leave one year out cross-validation. For SVR, R2 was 0.70, the Root Mean Square Error (RMSE) was 0.50 MT/ha and Mean Absolute Percentage Error (MAPE) was 27.6%. On the other hand for RF these were 0.69, 0.51 MT/ha and 29.3% respectively. These results are promising and should be tested in specific applications to understand if they are good enough for use.

A comparison of traditional and machine learning corn yield models using hyperspectral UAS and Landsat imagery

“…Several studies have focused on the exploitation of remote sensing data and indices for use in constructing mathematical models for forecasting maize silage yield. [5][6][7][8][9][10][11][12] Two examples of vegetation indices (VIs) that have shown promise as predictors are the normalized difference vegetative index (NDVI) and enhanced vegetative index (EVI). Both indices are regularly used in linear and exponential regression models:…”

“…where a and b are fitted parameters and V I represents Both linear and exponential models have been shown to vary in performance based on geographic region and scale, i.e., field vs county-wide yield estimates. [8][9][10] These models thus can safely rely on low spatial and spectral resolution data, i.e., typical multispectral systems, provided that there is high confidence in the collection system's radiometric calibration. The National Aeronautics and Space Administration's (NASA) Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) imagery systems, Planet Labs' Dove and Superdove constellations, and the European Space Agency's (ESA) Sentinel satellites provide calibrated reflectance imagery which have been used for such yield predictions.…”