In this manuscript, a new methodology based on a deep learning model using a Siamese network and attention module was proposed to classify crop cultivation areas, such as onion and garlic, from multitemporal PlanetScope images in South Korea. To consider the seasonal characteristics of crops in the model, training data were constructed from multitemporal satellite images. It was generated using PlanetScope satellite imagery from January and April, corresponding to the seasonal growth period of onion and garlic, in South Korea. Image patches were generated by considering the ratio of crops to minimize the influence of imbalanced data in the training process. Siamese FC-DenseNet with an attention module model (SFC-DenseNet-AM) is proposed, and the attention module is used to classify cultivated crop areas. Based on the proposed network, we extract cultivated crop areas using preliminary cultivation information. The results of the experiment using PlanetScope images indicate that image classification for cultivated areas was effectively performed using the proposed deep learning model. The model's performance, with F1-scores of 0.823 (garlic) and 0.774 (onion), was verified through an ablation study.
Seonkyeong Seong, Anjin Chang, Junsang Mo, Sangil Na, Hoyong Ahn, Jaehong Oh, Jaewan Choi, Crop classification in South Korea for multitemporal PlanetScope imagery using SFC-DenseNet-AM, International Journal of Applied Earth Observation and Geoinformation,Volume 126, 2024, 103619, ISSN 1569-8432, https://doi.org/10.1016/j.jag.2023.103619.