← Back to Portfolio
Agricultural Remote Sensing · Published 2023

Satellite-Based Crop Yield Prediction with Deep Learning

A deep-learning approach for predicting cabbage and radish yields in Gangwon-do using multi-temporal Landsat 8 observations and crop-related spectral and thermal indicators.

Published Deep Learning Crop Yield Landsat 8 Remote Sensing
Read Paper ↗
01 · Overview

Predicting crop yield from seasonal satellite observations

This study investigates whether multi-temporal Landsat 8 imagery can support regional yield prediction for cabbage and radish. Satellite observations were combined with vegetation and thermal indicators to represent crop conditions during the growing season.

Study period
2015–2020
Growing-season observations from June to September
Satellite
Landsat 8
Surface reflectance and thermal observations
Target crops
2 crops
Cabbage and radish in Gangwon-do
02 · Approach

Seasonal satellite features for CNN-based yield prediction

The growing season was divided into June–July and August–September. Eleven Landsat-based variables were summarized as histograms and used to train a CNN, with a reference CNN and Random Forest included for comparison.

01

Satellite preprocessing

Non-field areas and cloud-contaminated observations were removed before extracting crop-related information from each administrative region.

02

Temporal feature construction

Observations were grouped into two seasonal periods and represented as 32-bin histograms for each input variable.

03

Yield modeling

A CNN was trained to predict crop yield and evaluated against a reference CNN and Random Forest baseline.

Surface Reflectance NDVI EVI LAI LST CNN
Graphical abstract of cabbage and radish yield prediction
Study workflow. Landsat 8 observations and crop-yield statistics are processed into seasonal histogram features and evaluated with CNN and Random Forest models.
03 · Results

Crop-specific performance across the evaluated models

The proposed CNN achieved the strongest result for radish yield prediction. For cabbage, Random Forest outperformed the CNN, highlighting differences in model behavior across crop types.

Radish · Proposed CNN

Best result among evaluated models

0.705 Correlation r
1,553 RMSE kg/10a
Cabbage · Proposed CNN

Competitive, but below Random Forest

0.635 Correlation r
1,358 RMSE kg/10a

Random Forest achieved the strongest cabbage result (r = 0.684, RMSE = 1,163 kg/10a).

04 · Publication

Publication

Satellite-Based Cabbage and Radish Yield Prediction Using Deep Learning in Kangwon-do

Hyebin Park, Yejin Lee, Seonyoung Park
Korean Journal of Remote Sensing · Vol. 39, No. 5-3 · 1031–1042 · 2023