Optimizing time windows of Sentinel-2 images for deep learning-based cropland segmentation through iterative multi-criteria decision analysis

Precision agriculture depends on accurate cropland segmentation from satellite imagery. However, unoptimized satellite time series can lead to the inclusion of redundant information, increasing data volume and reducing the efficiency of deep learning (DL) model training. This study proposes a method that systematically employs iterative Multi-Criteria Decision Analysis (MCDA) to analyze the accuracy of DL models for cropland segmentation and to guide the selection of optimal time windows from Sentinel-2 time series imagery. Our method achieves a balance between segmentation accuracy and data usage. Optimized 3-month time windows can deliver overall accuracy of 90.9%, with only a 0.7% difference compared to using 12 months. The method was evaluated across multiple study areas characterized by diverse environmental conditions and distinct reference datasets. This demonstrates equal or superior results for optimized time windows compared to seven consecutive months in the growing season, underscoring the robustness and generalizability of the method.