An Approach Based on Web Scraping and Denoising Encoders to Curate Food Security Datasets

Abstract

Ensuring food security requires the publication of data in a timely manner, but often this information is not properly documented and evaluated. Therefore, the combination of databases from multiple sources is a common practice to curate the data and corroborate the results; however, this also results in incomplete cases. These tasks are often labor-intensive since they require a case-wise review to obtain the requested and completed information. To address these problems, an approach based on Selenium web-scraping software and the multiple imputation denoising autoencoders (MIDAS) algorithm is presented for a case study in Ecuador. The objective was to produce a multidimensional database, free of data gaps, with 72 species of food crops based on the data from 3 different open data web databases. This methodology resulted in an analysis-ready dataset with 43 parameters describing plant traits, nutritional composition, and planted areas of food crops, whose imputed data obtained an R-square of 0.84 for a control numerical parameter selected for validation. This enriched dataset was later clustered with K-means to report unprecedented insights into food crops cultivated in Ecuador. The methodology is useful for users who need to collect and curate data from different sources in a semi-automatic fashion.