IoT-based Bacillus Number Prediction in Smart Turmeric Farms Using Small Datasets

The Bacillus genus is one of the most commercially exploited bacteria in the agro-biotechnology industry, and the Bacillus information is very useful for crop growth. Most existing studies on the analysis of the amount of Bacillus were conducted in laboratories. Performing such a task on open field farming is difficult because only a small dataset is available during a long observation period for the soil analysis of Bacillus. For example, turmeric growth takes 9 months with one soil sample per month, and we found that increasing the frequency of soil analysis for turmeric growth is not practically useful. Therefore, we can only collect a very small dataset for AI training. This paper proposes the AgriTalk approach that predicts the amount of Bacillus based on novel IoT and machine learning technologies. AgriTalk uses a small dataset (5 data items) per farm for training and performs prediction for the subsequent 4 months. Good results are obtained. Specifically, the inference MAPEs (Mean Absolute Percentage Errors) range from 6.73% to 19.76%. In the experiments of five farm fields, we have correctly captured the trends for the number of changes of Bacillus. Such prediction provides useful information for fertilization management. Our prediction is more accurate for farms covered by peanut shells (the average MAPE is 13.24%) than for farms covered by rice husks (the average MAPE is 15.43%).