Prediction of Teff Yield Using a Machine Learning Approach

Teff is one of the main ingredients in everyday food for most Ethiopians, and its production mainly depends on natural conditions of the climate, unpredictable changes in the climate, and other growth factors. Teff production is extremely variable on different occasions and creates complex scenarios for prediction of yield. Traditional methods of prediction are incomplete and require field data collection, which is costly, with the result being poor prediction accuracy. Remotely sensed satellite image data has proven to be a reliable and real-time source of data for crop yield prediction; however, these data are enormous in size and difficult to interpret. Recently, machine-learning methods have been in use for processing satellite data, providing more accurate crop prediction results. However, these approaches are used in croplands covering vast areas or regions, requiring huge amounts of cropland mask data, which is not available in most developing countries, and may not provide accurate household level yield prediction. In this article, we proposed a machine learning based Teff Yield Prediction System for smaller cropland areas using publicly available multispectral satellite images, that represent spectral reflectance information related to the crop growth status collected from different satellites (Landsat-8, Sentinel-2). For this, we have prepared our own satellite image dataset for training. A Convolutional Neural Network was developed and trained to be fit for a regression task. A training loss of 3.3783 and a validation loss of 1.6212 were obtained; in other words, the model prediction accuracy was 98.38%. This shows that our model's performance is very promising.