Gross Paddy Yield Estimation Using a Lightweight Deep Learning Model

One of the important roles of a government is to ensure food security among its citizens, which is highlighted in the second Sustainable Development Goal - Zero Hunger. Therefore, the government needs to effectively monitor staple food production and distribution so that the supply always meets the demand of their citizen. In Asia, rice is the most important staple food that is produced by many countries but only a few of them managed to export the rice due to no surplus compared to the local consumption. In order to effectively monitor the paddy plantation, many government agencies have explored the usage of remote sensing imagery due to the main advantage of wide coverage areas with a low-cost repeated sample sensing platform. However, manually monitoring the paddy fields is a tedious process that needs to be automated, especially for the task of gross paddy yield estimation. Therefore, this paper proposed a novel lightweight deep regressor network to predict the gross paddy yield using Pleiades satellite imagery. In this study, a single day information of days after planting is used to model the yield prediction, which has been set at the age of 60 days. The ground truth has been manually annotated by the field officers, and organized in the MakGeoPadi system. The base of the regressor model is MobileNet V2, which has been enhanced by using optimal attention mechanism embedding. Five variants of the proposed model have been analyzed that differ in terms of the embedding placement and the set of convolutional filters used in the embedding. The best results with a mean accuracy of 0.715 is obtained when the embedding is placed at the second inverted residual block of the regressor model with a set of (32,64,128) convolutional filters used in the attention mechanism. The results show that the proposed model performance exceeds the other benchmarked models and is suitable to be implemented for predicting the paddy yield.