Automating Crop-Field Segmentation in High-Resolution Satellite Images: A U-Net Approach with Optimized Multitemporal Canny Edge Detection

Reliable and efficient crop field segmentation is a fundamental pre-requisite for statistical analyses of agricultural practices. Traditional methodologies such as the Canny-Watershed (CW) algorithm require expert tuning of parameters for optimal results. This paper introduces an innovative approach for crop field segmentation in high-resolution satellite images, leveraging the use of multi-temporal Canny edge detection to train convolutional neural networks (CNNs) and fully automate the segmentation process. The Canny filter, applied to Sentinel-2 multi-temporal data, provides refined input for training ResUnet models, facilitating the generation of a generalized training dataset. ResUnet allows the model to learn complex features from diverse data, encapsulating seasonal changes. The dataset was specifically designed to enable the model to make accurate predictions from a single image, significantly outperforming traditional Canny filter predictions. In addition, the ResUnet may be applied to multiple images, generating output masks that, when overlayed, produce better results with respect to the multi-temporal Canny approach, demonstrating superior ability in recognizing real field boundaries while reducing false detections.To enhance generalizability, ResUnet is trained on a varied global dataset capturing a wide range of agricultural conditions and seasonal variations. This generalized model is tested across different regions and seasons, and preliminary results indicate that the proposed approach offers operational efficiency and accuracy in automating crop field segmentation.