Core Concepts
The author reports benchmarking results of deep learning models on the MLCommons CloudMask Benchmark, highlighting the best model's performance and computational efficiency.
Abstract
This paper discusses the evaluation of deep learning models on the MLCommons CloudMask Benchmark, focusing on cloud masking in atmospheric sciences. The study presents benchmarking results, model comparisons, and computational performance metrics. Various methods, including rule-based and deep learning techniques, are analyzed for cloud masking tasks using satellite images from Sentinel-3. The study emphasizes accuracy, training time, and inference time as key evaluation criteria. The authors provide insights into the challenges and advancements in cloud masking algorithms using AI technologies.
Stats
Our benchmarking results include the highest accuracy achieved on the NYU system.
The dataset consists of 1070 images captured at day and night with dimensions 1200 x 1500.
The cloud masking benchmark contains 180 GB worth of satellite image data.
The U-Net model is used for image segmentation in cloud masking tasks.
The average accuracy achieved over five runs is 0.889.
Quotes
"MLCommons Science Working Group has developed four scientific benchmarks in varying fields."
"Deep learning techniques have shown superior performance in generating cloud masks."
"The reference implementation uses a U-Net model for image segmentation."