Predicting Canopy Height of Crops Using ICESat-2 Photons, SkySat Images and ResUNet Deep Learning Model

Le résumé fourni par la source

Crop canopy height is an important parameter in precision agriculture (PA). While contemporary studies have used deep learning models to predict canopy height, most of these studies target forest areas with less focus on croplands. Estimation errors of these models also tend to be greater and less effective is predicting canopy height of crops which tend to have lower height. This study predicted crop canopy height to a sub-meter vertical accuracy using the ResUNet deep learning model, Ice, Cloud and Land Elevation Satellite – 2 (ICESat-2) photons and high-resolution SkySat images. ICESat-2 height measurements and SkySat images were captured over croplands in the Coastal Bend Region of Texas, USA on July 05, 2023. The ResUNet model was used to establish a relationship between ICESat-2 absolute canopy height and the SkySat image, enabling the development of a wall-to-wall canopy height model (CHM) for the croplands. Evaluation metric (root mean square error - RMSE) confirmed the accuracy of the model (training RMSE = 0.35 m, test RMSE = 0.48 m, validation RMSE = 0.49 m), achieving sub-meter vertical accuracy. Comparison of the SkySat CHM with independent height data obtained from Uncrewed Aircraft Systems (UAS) through SfM-MVS (Structure-from-Motion – Multi-View Stereo) showed a moderate correlation (r = 0.55) and a mean overestimation of 20 cm. These results demonstrate the potential of integrating ICESat-2 and SkySat data with deep learning techniques for precise crop height estimation in support of PA.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.