Enhancing machine learning-based crop mapping with high-quality training samples

While machine learning (ML) approaches are able to produce crop maps through the classification of remotely sensed imagery, the acquisition of high-quality training samples for ML remains challenging. In this paper, we propose a sample evaluation scheme to address this issue. Firstly, an unsupervised ML is used to generate objective-based clusters, which serve as the basis for stratifications that reduce spatial redundancy in the sampling. Secondly, samples are randomly collected based on the stratification map, producing multiple sets of samples with varying size and spatial distribution. Lastly, the scheme evaluates the representativeness of individual samples by considering multiple features, as expressed by sample representativeness indicator, and introduces a comprehensive representativeness indicator (CRI) for each aggregated sample set. Based on this scheme, we hypothesize that the CRI can serve as a measure of the quality of a sample set. To test this hypothesis, we conducted a series of crop mapping experiments using support vector machine (SVM, a supervised ML) with different sample sets. Results show that: (1) There is an optimal sample size below which mapping accuracies vary significantly when different sample sets are employed. (2) When the sample size falls below the optimal threshold, choosing a sample set with a higher CRI robustly yields higher mapping accuracy. (3) Mapping accuracies and CRIs exhibit a significant correlation. These findings imply that the proposed sample evaluation scheme not only aids in collecting high-quality training samples for ML-based crop mapping but also showcases the capability to predict the accuracy of crop mapping by examining the inherent features of the collected samples.