Development Of A Self-learning Intelligent Machine For The Geographic Information System (GIS) To Classify And Diagnose Sugarcane Production From Satellite Images

Abstract

Sugarcane is an essential economic cultivation of Thailand, obtaining a GDP of approximately 200,000 million Baths annually. However, the sugarcane industry has many crises, especially sugarcane yield and production fluctuations. This project aims to develop a self-learning intelligence kit for a geographic information system to 1) classify the sugarcane area and 2) forecast sugarcane's yield. The location is in Sa Kaew Province. In the classification, Landsat 9 image and Sentinel-1A are acquired and associated with three AI algorithms i.e. artificial neural network, random forest, and gradient boosting. In yield forecasting, four vegetation indices extracting from Landsat 8 imagery are incorporated with five AI regressor algorithms i.e. multiple-linear regressor, multi-layer perceptron, decision tree regressor, random forest regressor, and gaussian process regressor.
Classification result reveals that the random forest provides the highest efficiency, attaining an accuracy of 0.76. The prediction of the sugarcane class reaches 0.89. Yield forecasting result reveals that the random forest regressor also provides the highest efficiency, achieving a coefficient of determination of 0.79. The mean average error reveals 0.79 ton/rai. In the application of AI, input data and hyperparameter testing should be considered for the highest efficiency of models.