Sampling design and dataset size for an accurate classification

Machine learning models applied to remote sensing (RS) imagery are a current trend. However, the sample design and the minimum dataset size are bottlenecks. This presentation aims to test whether a design with homogenous plots and buffer sections increases accuracy in mapping plastic mulched farmland (PMF). The comparison of three designs presented the impact of noise on model accuracy along the sample size variation. We tested k-nearest neighbors (kNN) and artificial neural networks (ANN) in a pixel-based binary classification. The kNN was more constant and sample-dependent, whereas the ANN showed unstable results with few samples. Statistically, the minimum sample size must be 1850, but kNN reached a stable balanced accuracy above 0.99 with 550 samples and ANN only with 2000. We provide insights into sampling techniques and model learning. The Z design is adequate for sampling small polygons in images with low or medium spatial resolution commonly found in RS.