Improvement in crop mapping from satellite image time series by effectively supervising deep neural networks

Deep learning methods have achieved promising results in crop mapping using satellite image time series. A challenge still remains on how to better learn discriminative feature representations detect types when the model is applied unseen data. To address this and reveal importance of proper supervision deep neural networks improving performance, we propose supervise intermediate layers a designed 3D Fully Convolutional Neural Network (FCN) by employing two middle methods: Cross-entropy loss Middle Supervision (CE-MidS) novel method, namely Supervised Contrastive (SupCon-MidS). This method pulls together features belonging same class embedding space, while pushing apart from different classes. We demonstrate that SupCon-MidS enhances discrimination clustering throughout network, thereby network performance. In addition, employ output methods, F1 Intersection Over Union (IOU) loss. Our experiments identifying corn, soybean, Other Landsat series U.S. corn belt show best set-up our IOU+SupCon-MidS, able outperform state-of-the-art mIOU scores 3.5% 0.5% average testing its accuracy across year (local test) regions (spatial test), respectively. Further, adding improves 1.2% 7.6% local spatial tests, conclude plays significant role The code data are available at: https://github.com/Sina-Mohammadi/CropSupervision.