Automatic Generation of Training Samples Based on Local Samples Phenological Information for Crop Type Classification

Reliable crop type classification is essential for ensuring food security. With advancements in computer technology, machine learning has become an indispensable tool in this domain. However, their successful implementation is often constrained by the scarcity of training samples. To address this challenge, this study proposes an effective method for automatically generating training samples using a limited number of localized ground-truth labels. The proposed method involves two steps: (1) automatically generating candidate samples by leveraging the phenological information from a small number of local samples within the study area using Time-Weighted Dynamic Time Warping (TWDTW) technology; (2) Cleaning the candidate samples using temporal NDVI features. We conducted multiple experiments in the study area located at the border of North Dakota and Minnesota. By combining samples collected from CDL data with generated samples, we demonstrated the effectiveness of the proposed method in enhancing classifier performance. The results indicated that, compared to using only the original samples, the average overall accuracy of RF and SVM classifications improved by 7.66% and 4.24%, respectively, when employing the generated training samples. This indicates that the generated training samples exhibit high reliability. This study provides a feasible approach for crop classification in regions with scarce ground data.