Addressing sample imbalance in deep learning based multi-temporal crop classification: a study in Jhansi, India

Accurate mapping of crop types is essential for addressing food security, crop inventory and supporting farmers in decision making to improve production and manage agricultural practices. Generally, crop mapping studies either ignore minority crops or aggregate them into ‘other crops’ class during classification. This study addresses the challenge of class imbalance in crop type classification using remote sensing imagery, focusing on agricultural fields in Jhansi District, Uttar Pradesh, India, during the Rabi cropping season. A one-dimensional convolution neural network (1D CNN) based sequential deep learning model was employed using multitemporal Sentinel 2 imagery. Algorithm level and data level balancing techniques were examined. Categorical cross-entropy, focal loss, and class weighted loss functions were compared to reduce class imbalance, with the latter proving most effective. Additionally, data level balancing methods including undersampling, oversampling, and hybrid approaches were investigated, with undersampling outperforming oversampling techniques. Evaluation metrics such as Overall Accuracy, F1-score, Precision, Recall and G-Mean score were employed, with the G-Mean score preferred for assessing individual class accuracy in imbalanced datasets. Findings indicate that algorithm level balancing with class weighted loss shows the best results, improving overall accuracy to 67.50% with a G-Mean score of 59.46%. This research advances crop mapping methodologies by providing insights into effective strategies for addressing class imbalance in deep learning–based classification, which is crucial for accurate classification of crops.