Performance Evaluation of Satellite Temporal Aggregation Methods and Sample Size Training Data for Crop Mapping

The generation of agricultural land cover maps through remote sensing has expanded in recent years. Having accurate and nearly real-time maps is crucial for decision- makers. However, there are significant challenges in terms of the data required to produce these maps. Firstly, it is necessary to select and summarize the most useful satellite imagery. Secondly, collecting ground reference data from the campaign in the study area is costly in terms of money, time and human resources. The aim of this study was to evaluate and compare the performance of three methods for temporal aggregation of satellite imagery, while extending the training datasets used to map summer agricultural land cover. To achieve this, Sentinel-2 image mosaics were computed using three temporal aggregation methods: one selecting the images with the least cloud cover, another calculating median values, and the third retrieving data related to maximum NDVI values. The ground reference dataset was divided into two independent sets: one for training and the other for validation. To analyze the effect of training dataset size, successive tests were conducted with increasing dimensions of the training dataset. The tests were carried out using a gradient boosting classification model in a purely agricultural area in the province of Cordoba, covering the summer agricultural campaigns of 2015/16, 2016/17, and 2017/18. The results showed that the composites produced by the median and maximum NDVI methods had the highest statistical metrics for assessing the accuracy of thematic maps. For the studied area and seasons, a training set size of 65 samples (plots) per crop class allowed achieving accuracy values of 90%.