Keywords
Self organizing map; unsupervised learning; agronomic variables; transitions
Start Date
6-7-2022 7:40 AM
End Date
6-7-2022 8:00 AM
Abstract
Spatial and temporal changes in Agricultural Land Systems (ALS) are often interlinked and coexist within quite large areas determining a complex set of structures and transitions. To detect these changes is necessary to reduce the complexity of the input information to few typologies which can be easier managed and interpreted. In the recent period, the development of data driven methods based on machine learning enlarged the availability of sophisticated techniques that allow researchers to discover pattern in environmental dataset. Among them, Self-Organizing Map (SOM), an unsupervised competitive learning neural network, performs particularly well to identify clusters from a multivariate dataset. This study aims to identify and characterize the main ALS in the Western Mediterranean area and to compare their pattern distribution during a ten-year time span. Input data came from the official national agricultural censuses of Spain, Portugal, France, and Italy, collected at municipality scale. To reach our objective, we developed a two-step unsupervised clustering procedure, based on SOM for dimensionality reduction and followed by hierarchical clustering to perform the final partitioning, where similar units are aggregated into larger clusters. These final clusters (five at each census-period) have been investigated in terms of distribution of the agronomic variables. This allowed to detect and characterize the predominant ASL as well as their dynamic, evaluated as transitions among clusters revealed in the two census-periods. At the end of the entire process, results have been mapped under a GIS environment, allowing to visualize the revealed patterns over the geographic space.
Change dynamics of agricultural land systems in Europe: an innovative approach based on an unsupervised clustering technique
Spatial and temporal changes in Agricultural Land Systems (ALS) are often interlinked and coexist within quite large areas determining a complex set of structures and transitions. To detect these changes is necessary to reduce the complexity of the input information to few typologies which can be easier managed and interpreted. In the recent period, the development of data driven methods based on machine learning enlarged the availability of sophisticated techniques that allow researchers to discover pattern in environmental dataset. Among them, Self-Organizing Map (SOM), an unsupervised competitive learning neural network, performs particularly well to identify clusters from a multivariate dataset. This study aims to identify and characterize the main ALS in the Western Mediterranean area and to compare their pattern distribution during a ten-year time span. Input data came from the official national agricultural censuses of Spain, Portugal, France, and Italy, collected at municipality scale. To reach our objective, we developed a two-step unsupervised clustering procedure, based on SOM for dimensionality reduction and followed by hierarchical clustering to perform the final partitioning, where similar units are aggregated into larger clusters. These final clusters (five at each census-period) have been investigated in terms of distribution of the agronomic variables. This allowed to detect and characterize the predominant ASL as well as their dynamic, evaluated as transitions among clusters revealed in the two census-periods. At the end of the entire process, results have been mapped under a GIS environment, allowing to visualize the revealed patterns over the geographic space.
Stream and Session
false