How AI Predicts Short-Term Rainfall from Radar Images: A Competition Walkthrough

Competition Overview

CIKM AnalytiCup 2017, organized by the ACM CIKM conference and co‑hosted by the Shenzhen Meteorological Bureau and Alibaba, challenged participants to improve short‑term precipitation forecasts using radar image sequences.

Data Description

The dataset provides 10,000 samples, each containing 60 radar images captured over the past 90 minutes (6‑minute intervals, 15 frames) at four altitudes (0.5 km, 1.5 km, 2.5 km, 3.5 km). Each image is a 101 × 101 grid covering a 101 km × 101 km area, with pixel values representing the transformed radar reflectivity factor Z.

Task Objective

Predict the total ground precipitation at the central coordinate [50, 50] for the next 1–2 hours, using mean‑square‑error as the loss function.

Algorithm Architecture

The solution consists of three stages: preprocessing, feature extraction, and model training. Preprocessing stitches overlapping local radar patches into a larger image and uses SIFT descriptors to estimate motion vectors. Feature extraction creates three groups of descriptors: spatio‑temporal vectors, statistical summaries of cloud shape, and extrapolated radar reflectivity images for the target location. The main model is a convolutional neural network (CNN) with two convolution‑pooling layers; the flattened image features are concatenated with non‑image attributes and fed into two fully‑connected hidden layers.

Image Stitching

Local radar patches overlap; template matching aligns them to reconstruct a wider spatial view, which extends the trajectory tracking range and reveals the global cloud structure.

Trajectory Tracking

Based on the Taylor frozen hypothesis, cloud motion is approximated as a translation with the local average convection velocity U. SIFT matches across frames provide pixel‑level displacements δx, from which U is derived for each grid point.

Feature Extraction

Three feature groups are built:

Time‑extrapolated reflectivity images obtained by projecting cloud positions forward using the estimated velocity field.

Spatio‑temporal statistical descriptors (mean, max, variance, count of extreme values) computed for each altitude.

Global cloud‑shape descriptors, including reflectivity histograms, motion vectors, acceleration, streamline curvature, and SIFT‑histogram statistics.

Model Training

The CNN receives a 41 × 41 patch centered on the target location (covering three altitude layers) as image input. Two convolution‑pooling layers are followed by flattening, concatenation with the non‑image features, and two hidden layers. Dropout with keep‑probability 0.65 prevents over‑fitting, and the Adam optimizer converges after roughly 1,200 iterations.

Conclusion

The approach avoids generic ImageNet models such as InceptionNet or ResNet, instead combining traditional SIFT‑based motion tracking with a purpose‑built CNN to capture the spatio‑temporal dynamics specific to meteorological radar data.