Satellite-Based Detection of Farmland Manuring Using Machine Learning Approaches

In agriculture, manuring offers several benefits, which include improving soil fertility by adding essential nutrients like nitrogen, phosphorus, and potassium. It enhances soil structure, promoting better water retention and aeration, which positively impacts plant growth. Additionally, manuring can suppress certain plant diseases and pests, contributing to healthier crops. However, improper handling and application of manure can pose risks, such as the spread of pathogens and water pollution. To mitigate these risks, it is crucial to follow proper storage and composting practices, but also to observe correct application periods and techniques. Spaceborne Earth observation can contribute to monitor manuring, which helps mapping possible derived risks; however, manure detection from satellite data is still an open problem. The aim of this research work is an automated, Machine Learning (ML-)based approach, to detecting manure application on crop fields in time sequences of space-borne, multi-spectral optical Earth Observation data. Among a group of different spectral indexes extracted from multispectral satellite acquisitions, those most impacted by manure application have been identified, and used to train and test various Machine Learning models. This led to discover that the spectral signature of manure application learnt from the concerned training data cannot be easily transferred to different contexts. On the other hand, integrating thermal data allows to improve accuracy, despite possible thermal to multi-spectral sampling mismatch in time series. The addition of radar data offered instead no significant contribution to system performances. The identified method is the first step towards large-scale, consistent monitoring of manure application to check compliance with environmental regulations.