Dealing With Clouds and Seasonal Changes for Center Pivot Irrigation Systems Detection Using Instance Segmentation in Sentinel-2 Time Series

The automatic detection of Center Pivot Irrigation Systems (CPIS) is fundamental for establishing public policies, especially in countries with a growth perspective this technology, like Brazil. Previous studies to detect CPIS using deep learning used single-date optical images, containing limitations due seasonal changes and cloud cover. Therefore, research aimed Sentinel-2 multitemporal images (containing six dates) instance segmentation, considering variations different proportions cloudy generalizing the models diverse situations. We novel augmentation strategy, which, each iteration, were randomly selected from time series (from total 11 random order. evaluated Mask-RCNN model ResNext-101 backbone COCO metrics on testing sets ratios cloudless ( $< 20\%$ ) notation="LaTeX">$>75\%$ ), zero (6:0) up one image five (1:5). found that provided best [80% average precision (AP), 93% AP 0.5 intersection over union threshold (AP50)]. However, results similar (74% AP, 88% AP50) even extreme scenarios abundant presence (1:5 ratio). Our method provides more adaptive way map series, significantly reducing interference such as cover, atmospheric effects, shadow, missing data, lack contrast surrounding vegetation.