CROP YIELD PREDICTION IN NEPAL

Abstract

Agriculture being one of the major economic activities in Nepal, estimating the crop yield plays an important role in decision making and food security. While the traditional agricultural systems mostly depend on ground-survey data, the freely available historical data of crops and climate can be excellent tool to support these systems by estimating the crop yields before harvest. In this work, a dataset containing 40 years of crop and monthly-climate data of 15 districts of Nepal was prepared and then augmented with 28 years of crop and monthly-climate data of 25 districts of India. Paddy, wheat, maize, oilseed, sugarcane, potato, lentil, and black gram are the crops and temperature, humidity, windspeed, precipitation, and soil wetness are the input parameters used for data collection. Traditional ML algorithms like Linear regression (LR), Support Vector Regressor (SVR), K-Nearest Neighbor (KNN), Decision Tree Regressor (DT) and Deep neural network (DNN) along with hybrid models like proposed Stacking Regression (SR) and proposed CNN-DNN model were trained on the Nepali dataset and augmented dataset. The proposed solution that converts monthly agroclimatic tabular data into image and then fed to hybrid CNN-DNN model was unable to outperform DNN. CNN-DNN, SR and DNN produced 17-38%, 13-28% and 10-29% mape respectively for different crops at best.