Evaluation Of Machine Learning Classification Methods For Rice Detection Using Earth Observation Data: Case Of Senegal

Abstract

Agriculture is considered one of the most vulnerable sectors to climate change. In addition to rainfed agriculture, irrigated crops such as rice have been developed in recent decades along the Senegal River. This new crop requires reliable information and monitoring systems. Remote sensing data have proven to be very useful for mapping and monitoring rice fields. In this study, a rice classification system based on machine learning to recognize and categorize rice is proposed. Physical interpretations of rice with other land cover classes in relation to the spectral signature should identify the optimal periods for mapping rice plots using three machine learning methods including Support Vector Machine (SVM), Random Forest (RF), and Classification and Regression Trees (CART). The database is composed of field data collected by GPS and high spatial resolution (10 to 30 m) satellite data acquired between January and May 2018. The analysis of the spectral signature of different land cover show that the ability to differentiate rice from other classes depends on the level of rice development. The results show the efficiency of the three classification algorithms with overall accuracies and Kappa coefficients for SVM (96.2%, 94.5%), for CART (97.6%, 96.5%) and for RF (98% 97.1%) respectively. Unmixing analysis was made to verify the classification and compare the accuracy of these three algorithms according to their performance.