Training sample selection for robust multi-year within-season crop classification using machine learning

Within-season crop classification using multispectral imagery is an effective way to generate timely crop maps that can support water and crop management; however, developing such models is challenging due to limited satellite imagery and ground truth data available during the season. This study investigated ways to optimize the use of multi-year samples in a within-season crop classification model, aiming to enable accurate within-season crop mapping across years. Our study focused on classifying field-scale corn/maize, cotton, and rice in south-eastern Australia from 2013 to 2019. The crop classification model was based on the random forest and sup-port vector machine algorithms applied to Landsat 8 multispectral bands. We designed four experiments to understand the influences of training sample selection on model accuracy. Specifically, we analyzed how the within-season classification accuracies are affected by 1) training sample size; 2) proportions of classification classes; 3) the inclusion of a non-crop class (e.g., fallow land) in the training sample, and 4) training samples collected from different years. We found that 1) the training sample size should be sufficiently large to ensure within-season classification accuracy; 2) using training samples for each crop type in proportion to their occurrence within the landscape results in more accurate multi-year classification; 3) the inclusion of the non -crop class can reduce the accuracy with which crop types are distinguished, so the proportion of the non-crop class should be maintained at a relatively low level, and 4) predicting the current year with training samples from previous years can lead to a minor decline in accuracy compared to using samples only from the current year. These training sample settings were adopted to develop a final model. We found that the model accuracy continues to improve as more input imagery is added as the cropping season progresses, with a rapid rate of initial improvement which then slows. December, the third month of the summer growing season, is the earliest time that reliable maps were generated, with an overall accuracy of 86 % and user's accuracies for all crops exceeding 80 %. Our proposed experiments are robust and transferable to other regions and seasons to assist the development of within-season crop maps, and can thus be valuable tools to support agricultural management.