The frequency of floods during the monsoon in South Asia has been increased over the years that raising the risk of loss and damage to rice crops. The staple food for majority of people in the world, resulted in production uncertainty. Timely and accurate loss and damage estimation is critical for effective compensation, relief distribution, crop insurance policy development, and ensure the food security in the country or in region. To quantify crop losses, the data on flood intensity is required, which eventually measures the extent of damage caused by floods in rice crops. Synthetic Aperture Radar (SAR)-based flood mapping methods have been validated and well accepted, but challenges remain in accurately mapping flood intensity. We uses a U-Net deep learning model with ResNet 34 backone for automatically mapping flood intensity using SAR and optical images, along with topographic information. The deep learning approach requires a large amounts of training dataset. We prepared them considering the different flood events in an ex-post flood case of the Ganga River basin in Nepal, India and Bangladesh from 2019 to 2024. The flood extension and intensity were labelled using the visual interaction in the Sentinel 1 SAR, Sentinel 2 and PlanetScope dataset where available. The duration of standing water and vegetation condition were major indicator to confirm the damage level manually. The digitized flood intensity level were refined with the field visit and validated with the stakeholder discussion. The extent of crop damage is affected by flood duration (i.e., how long standing water remains in the field) and the speed of the water current. Therefore, we used these dataset to fit the model. The U-Net model was trained with more than 20,000, 64 × 64 patches across three countries, covering different agro-ecological zones and topographic conditions in South Asia. The model was able to precisely map the flood intensity in different ecological zones and topography. The comparative analysis between the ground truth data with the model generated damage intensity showed high agreement: 94% over all accuracy and iou is 0.906 with validation loss is 0.03 demonstrates the capacity of the model. The study demonstrated, for the first time, the potential of using automated method that combined deep learning algorithm and remote sensing data for flood intensity mapping in the rice crop.