Using Deep Learning-based Super-Resolution to Improve Maps of Smallholder Crop Field Boundaries
Topics: Agricultural Geography
, Remote Sensing
, Africa
Keywords: agriculture, super-resolution, deep learning, field segmentation
Session Type: Virtual Paper
Day: Thursday
Session Start / End Time: 4/8/2021 03:05 PM (Pacific Time (US & Canada)) - 4/8/2021 04:20 PM (Pacific Time (US & Canada))
Room: Virtual 13
Authors:
Yunzhe Zhu, Clark University
Lyndon Estes, Clark University
,
,
,
,
,
,
,
,
Abstract
Accurate field mapping is critical for production estimation, agronomic practices, and land management. Smallholder farming, which is the dominant agricultural activity in Africa, is characterized by small field sizes and highly variable appearance within satellite images. Precisely mapping these fields requires high spatial-resolution and cloud-free satellite images. Super Resolution (SR) is a computer-vision technique for reconstructing high-resolution (HR) images from relatively low-resolution (LR) raw images. The deep learning-based SR directly learns the end-to-end mapping function from the LR image to the HR image through the neural network. To improve smallholder field mappings, we applied an SR procedure to generate 3 m resolution from 10 m Sentinel-2 multi-spectral imagery, using PlanetScope data as the source of training imagery. We first developed seasonal image composites to produce cloud-free imagery over a large scale. We then compared the capabilities of several deep learning models, such as CNN, ResNet, RED, and GAN, to produce SR images, and created an ensemble neural network architecture from these to improve the quality of SR. Finally, we implemented an existing field segmentation algorithm on the downscaled imagery and evaluated its performance relative to the same model applied to original PlanetScope data. The result provides insight into the ability of super-resolution approaches for improving the accuracy and performance of field segmentation algorithms.