Efficacy of machine learning and LiDAR data for crop type mapping

By Atman Prins

Accurate crop type maps are important for obtaining agricultural statistics such as water use or harvest estimations. The traditional approach to obtaining maps of cultivated fields is by manually digitising the fields from satellite or aerial imagery. However, manual digitising is time-consuming, expensive and subject to human error. Automated remote sensing methods have been a popular alternative for crop type map creation, with machine learning classification algorithms gaining popularity for classifying crop types from satellite imagery. However, using light detection and ranging (LiDAR) data for crop type mapping has not been widely researched.

This study evaluates the use of LiDAR data and machine learning algorithms for mapping vineyards. Vineyards are planted in rows spaced at various distances, which can cause spectral mixing within individual pixels and complicate image classification. Four resolution where used for generating normalized digital surface model and intensity derivatives from the LiDAR data. In addition, texture measures with window sizes of 3×3 and 5×5 were generated from the LiDAR derivatives. The different combinations of the resolutions and window sizes resulted in eight data sets that were used as input to 11 machine learning algorithms. A larger window size was found to improve the overall accuracy for all the classifier–resolution combinations. The results showed that random forest with texture measures generated at a 5×5 window size outperformed the other experiments, regardless of the resolution used. We conclude that the random forest algorithm used on LiDAR derivatives with a resolution of 1.5m and a window size of 5×5 is the recommend configuration for vineyard mapping using LiDAR data.