Design and implementation of an efficient and cost effective deep feature learning model for rice yield mapping

Abstract

Crop yield prediction before harvest is essential to address the instability of crop prices and ensure food security. Existing approaches of crop yield forecasting focus on survey data and are expensive. Remote sensing-based crop yield forecasting is a promising approach, especially in areas where field data is scarce. Recent studies used machine learning and deep learning techniques used modern representation learning ideas instead of traditionally used features that discarded many spectral bands available from the satellite imagery. A deep feature learning model using convolutional LSTM cells is used for forecasting rice yield from remote sensing satellite imagery. Convolutional LSTM with convolutional input and recurrent transformations directly captures spatial and temporal features of the input data. Feature selection is performed using principal component analysis to reduce the dimension of input data without much loss in the performance. Results suggest that features learned are highly informative and our proposed model performed better than other existing techniques.