Efficient & Sustainable Farming through Artificial Intelligence

One of today’s greatest challenges is feeding the world’s growing population sustainably. It is crucial to have efficient and sustainable farms that can maximize crop yield with minimum environmental impact and resource consumption, such as fertilizer and water usage. To achieve this goal, one critical and arguably the most important factor is to have an effective soil microbiome. The soil microbiome is the complex community of microorganisms inhabiting the soil that play many vital roles in the ecosystem. An effective microbiome can offer benefits to its host, including plant growth promotion, nutrient use efficiency, and control of pests and phytopathogens. For example, beneficial bacteria, such as Rhizobiales, can help plants capture nitrogen and other nutrients, improve soil fertility, and provide disease protection. On the other hand, harmful bacteria, such as Xanthomonas, can cause plant diseases, produce toxins, etc.

Based on this common knowledge, we further hypothesize that the composition of a soil microbiome can be a good indicator of crop yield. If this is true, information on soil microbiomes can be used to help farmers plan out how to allocate their resources and grow food in a more sustainable fashion.

In this work, using the data reported in [1], we tested different machine learning (ML) algorithms to identify the correlation between the microbiome composition and crop yield, as well as to identify the key microbes that contribute to crop yield. The dataset included soil samples with microbial counts of 630 different species from 1,344 plots, together with sweet corn yields on those plots. Four ML algorithms were investigated: Multiple Linear Regressor, Multi-layer Perceptron Regressor, K-Neighbors Regressor, and Decision Tree Regressor. Among them, Multiple Linear Regressor is a linear method, while the rest are nonlinear methods. The model input is microbiome composition or bacteria count, and the model output is the crop yield (measured in kilograms of crops per hectare of land area). Three different metrics were utilized to evaluate the prediction performance of different methods: coefficient of determination (R²), mean absolute error (MAE), and relative squared error (RSE). We found that a Decision Tree Regressor (DTR) model was the most accurate model, which delivered the highest R² and the smallest MAE and RSE.

This result confirms that soil microbiome composition can serve as a very good indicator for crop yield, and a good ML model can serve as the linkage. The ML model can provide guidance on resource allocation and farming management practices, for example, using biochar to increase the abundance of the microbes that have positive impacts on crop yield.

References:

[1] S. Nielsen et al., “Comparative analysis of the microbial communities in agricultural soil amended with enhanced biochars or traditional fertilisers,” Agric Ecosyst Environ, vol. 191, pp. 73–82, 2014.