A new approach to classify sugarcane fields based on association rules

Abstract

In order to corroborate the acquired knowledge of the human expert with the use of computational systems in the context of agrocomputing, this work presents a novel classification method for mining agrometeorological remote sensing data and its implementation to identify sugarcane fields, by analyzing Normalized Difference Vegetation Index (NDVI) series. The proposed method, called RAMiner (R ule-based A ssociative classifier Miner) creates a learning model from sets of mined association rules and employs the rules to constructs an associative classifier. RAMiner was proposed to deal with low spatial resolution image datasets, provided by two sensors/satellites (AVHRR/ NOAA and MODIS/Terra). The proposal employs a two-ways classification step for the new data: Considers the conviction value and the conviction-based probability (a weighted accuracy formulated in this work). The results given were compared with others delivered by well-known classifiers, such as C4.5, zeroR, OneR, Naive Bayes, Random Forest and Support Vector Machine (SVM). RAMiner presented the highest accuracy (83.4%), attesting it is well-suited to mine remote sensing data.