Application of Machine Learning Algorithms in Early Maize Yield Estimation Using Multiple Source Data in Kenya

Abstract

Agriculture and its interrelated activities form the main livelihood for most sub-Saharan Countries. Despite the importance of the sector, it continues to face several shortfalls that have led to cyclic food and nutrition insecurity over time. Ensuring food and nutrition security has been a common discussion agenda in the developing countries. The world population is estimated to reach 8.0 and 9.0 billion people by the year 2023 and 2050 respectively. This calls for supply of save food for all and an urgent need to increase the productivity of major food crops. To address the challenge of food and nutrition insecurity, it is important to involve and incorporate early planning options for the detection of any chances of depressed crop yields. Maize crop is a key food commodity in Kenya and it is almost equated to mean food security, with its shortage pointing to mean serious consequences. Maize yield prediction has faced several challenges with the information of how much is expected to be harvested coming too late to avert any serious consequences. Traditional methods for yields prediction have proved ineffective and inefficient in matching the rate of evolving challenges in yields forecasting. The typical methods for estimating maize yields are either, costly, presenting logistical nightmare besides providing prediction too late for any meaningful alternative options to be taken. Typical maize yield prediction over the years have not offered the expected results. The emergency of robust, efficient, innovative and effective crop yields prediction models offers great opportunity to mitigate the ongoing challenges. This study applied Machine Learning (ML) algorithms in the early maize yield estimation using multiple source datasets. Specifically, the study determined the most significant Vegetative Indices (VIs) in the prediction of early maize yields in Kenya, prescribed an appropriate Machine Learning (ML) algorithm for predicting early maize yields and established the optimal settings for early maize yields prediction in Kenya. This was achieved by utilizing free satellite imagery datasets using the Google Earth Engine (GEE) platform using cloud computing facility. Google Earth Engine (GEE) has immensely changed the way in which earth observations datasets are processed, analyzed and managed. It offers faster and more efficient way compared to other models. Two satellite datasets from Sentinel-2 and Landsat-9 were used to obtain VIs from geo-referenced smallholder farms from 2022 and 2023. Survey data was provided from State Department for Agriculture (SDA), Kenya Cereal Enhancement Programme – Climate Resilience Agricultural Livelihood (KCEP-CRAL) project. The study data was obtained from 171 counties in which various longitudinal surveys had been ongoing. From the accessed satellite imagery, 6 Vegetative Indices (VIs) where used namely, Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Difference Water Index (NDWI), Optimized Soil Adjusted Vegetation Index (OSAVI), Infrared Percentage Vegetation Index (IPVI) and Green Normalized Difference Vegetation Index (GNDI) from the different bands. Weather related variable obtained along the VIs included the monthly rainfall and Soil Surface Moisture (SSM). The data was cleaned to obtain cloud free images. Data mining was done using python scripts and analysis by RStudio. Feature extraction was undertaken using Principal Component Analysis (PCA) approaches to reduce factors, in which case, GNDVI (ɼ=0.05) and EVI (ɼ=0.02) for sentinel-2 were identified as less important while for landsat-9 imagery most variables were less important for model building. Random Forest (RF) algorithm was determined as the most evaluated appropriate Machine Learning (ML) algorithm (RMSE training set =3, RMSE test set = 5.4) compared to Multiple Regression (ML) and Support Vector Machine (SVM). The study established that sentinel-2 dataset provided better imagery under RF algorithm ML for estimating early maize yields in Kenya. The study recommends Sentinel-2 dataset in the maize yields estimation in Kenya that offers readily freely, available and efficient imagery. Investment into powerful capacity building on cloud computing and management of bigdata are critical considerable for its optimal settings. 1 including 2 control counties.