A Rainfall prediction model using long short-term neural networks for improved crop productivity: a case of maize planting in Machakos County

Wangome, Brian Mwathi
Journal Title
Journal ISSN
Volume Title
Strathmore University
Climate variability is a factor that affects crop productivity in Kenya. The unpredictable nature of weather patterns during the traditional long and short rain seasons has resulted in the rains starting earlier or later than expected. This unpredictability results in rainfed agriculture farmers experiencing losses on capital, fertilizers, and labor input and consequently declined agricultural productivity. The decline in food production also poses an existential threat to our nation’s food security and farmers’ incomes. Weather forecasts are aimed are reducing this uncertainty however, the sparse distribution of synoptic weather stations in Kenya that collect and monitor surface level meteorological conditions makes it hard for the Kenya Meteorological Department to guarantee a high spatial and temporal resolution. Therefore, the current forecast data disseminated to farmers is ‘coarse’, at the county and town level, which is of less significance to the smallholder farmer since this data does not factor in the topographical nuances within locations. The format of the weather forecasts is also technical for the farmers hence they resort to traditional methods in terms of planning for planting. The study proposed the use of deep learning techniques to build a rainfall forecasting model that accepted historical weather data and returned forecasted rainfall values in millimeters. The historical weather data was satellite data sourced from NASA’s Modern-Era Retrospective Analysis for Research and Applications Version 2 (MERRA-2). The historical data was used to train a Long Short-Term Memory neural network. An experimental approach was used to determine the number of epochs used in training the model and the number of timesteps/days into the future in which the most optimal model would forecast. In this study, the model forecasts 30 days into the future by looking at the past 60 days observed. The 30-day prediction model had a Root Mean Squared Error of 2.45 millimeters. Therefore, given the farmer’s Global Positioning System coordinates, the system can fetch past 60-day weather data and forecast the rainfall for the coming 30 days to help farmers to determine when to sow.
A Thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Information Technology at Strathmore University