A method to estimate densities of (Schonherr) adults captured in electronic probe traps in paddy based on deep neural networks.

Infestation by stored-grain insects causes qualitative and quantitative losses in stored grain. Effective estimation of insect densities is essential to minimize damage to stored grains. Probe trapping is a generally accepted method for detecting insect presence in bulk grains, but estimating insect densities based on trap counts has always been challenging. In this study, a novel method based on neural networks was proposed to assess the density of Cryptolestes pusillus (Schonherr) adults captured in electronic probe traps in paddy bulks. The data collection experiments were conducted in five bins (three filled with paddy at 10.7% and two at 14.0% moisture content) by deploying 15 electronic probe traps in each bin under insect densities of 0.1, 1.0, and 5.0 adults/kg. This study combined the trap count data and three key impact factors: grain temperature, moisture content, and trap location. We proposed a deep belief network for insect density estimation (DBN_IDE) that handled the feature extraction and data fusion of trap counts and the three key impact factors. The performance of DBN_IDE was evaluated by ablative and comparative experiments, which showed that DBN_IDE succeeded in efficiently combining trap counts and the three key impact factors using neural networks to estimate insect densities. DBN_IDE achieved an accuracy of 84.9% for insect density estimation. In addition, the optimal number of traps for density estimation was discussed, using five or six probe traps yielded optimal assessment results for the experimental setup of this study. Compared with the manual sampling method, the proposed method achieved better results. These results indicated that the proposed method for estimating the densities of adults captured in electronic probe traps in paddy bulks based on neural networks might be feasible for estimating insect densities during storage.