Abstract: Accurate crop type classification from remote sensing imagery plays a vital role in agricultural monitoring and precision farming, and its reliability critically depends on the representativeness of training samples, especially at the county scale where sample acquisition is costly and sample size is often limited. Traditional sampling strategies, such as random sampling, systematic (grid) sampling, or empirical stratification, primarily emphasize spatial uniformity or prior zoning, but often fail to adequately capture crop spectral variability and its interaction with sample size. To address this issue, this study takes county-level crop classification as an experimental scenario and develops a geographical similarity-based sampling strategy aimed at improving sample representativeness in feature space. Stratified random sampling and systematic sampling are adopted as baseline methods. Three classification models—support vector machine (SVM), random forest (RF), and temporal convolutional network (TCN)—are employed to conduct comparative experiments under multiple sample-size conditions, and model performance is evaluated in terms of classification accuracy and sample representativeness. Experimental results show that the advantages of sampling strategies are most pronounced under small-sample conditions. The similarity-based sampling strategy achieves effective coverage of the feature space with fewer samples, leading to noticeably higher accuracy than the baseline methods. Specifically, classification accuracy improves by approximately 2-12% in the SVM and TCN models, while differences among sampling strategies in the RF model remain relatively small (0-3%). As sample size increases, accuracy differences among the sampling strategies gradually diminish, indicating that the effectiveness of similarity-based sampling is most evident in sample-limited scenarios. Further analysis reveals that crop-specific spectral heterogeneity strongly influences sampling effectiveness. Crops with distinctive and stable spectral signatures can achieve high classification accuracy with limited samples, whereas crops with higher spectral heterogeneity or mixed spectral behavior benefit more from similarity-based sampling, which enhances sample representativeness by expanding coverage of the feature space. These results demonstrate that similarity-based sampling is particularly suitable for complex crop classification tasks under limited sampling conditions. The study area is a relatively flat agricultural plain with low environmental heterogeneity, which may constrain the full potential of similarity-based sampling under more complex environmental gradients. Therefore, its applicability to mountainous or highly heterogeneous regions requires further investigation. In addition, this study primarily uses classification accuracy as an indirect indicator of sample representativeness and does not explicitly model the quantitative relationship between representativeness and environmental complexity. Future work will focus on incorporating explicit representativeness metrics and exploring the coupling mechanisms among sample representativeness, landscape heterogeneity, and classification accuracy to further improve the robustness and theoretical foundation of sampling strategies for remote sensing–based crop classification.