Inversion of the high-resolution mechanically harvested cotton defoliation effect using GAN-based super-resolution reconstructed RGB satellite images

Abstract

The low spatial resolution of satellite remote sensing has become a major limiting factor in monitoring crop growth. Existing studies have effectively enhanced the resolution of remote-sensing imagery through super-resolution reconstruction (SR) techniques. Although these methods demonstrate good generalizability across different scenarios, they do not adequately address the anisotropic spatial structural characteristics of agricultural fields, particularly the influence of crop ridge orientation distribution on reconstruction outcomes. To address this issue, this study developed a super-resolution direction-aware generative adversarial network (SRDGAN) that incorporates multi-attention mechanisms and orientation-aware convolutions specifically designed for RGB satellite imagery of cotton farmlands. Large-scale monitoring models for defoliation and boll-opening rates were developed using feature selection and machine-learning techniques. The key findings include the following: (1) Significant correlations exist between features (vegetation indices, color components, and texture features) extracted from RGB satellite images and both defoliation and boll-opening rates. (2) A comparative analysis of SR methods revealed a progressive improvement in accuracy in the order Bicubic < EDSR< HAT< SRGAN < SRDGAN. (3) The RFE-XGBoost model achieved the optimal defoliation-rate monitoring accuracy using SRDGAN-enhanced imagery, whereas the RF-XGBoost model achieved the optimal boll-opening-rate monitoring accuracy. This study innovatively addresses the inherent conflict between spatial resolution and coverage range by incorporating directional perception into traditional GAN frameworks. The proposed methodology enables high-spatial resolution and large-scale precision management of cotton defoliation effects, providing crucial technical support for ensuring high-quality and efficient mechanical cotton harvesting in Xinjiang's modern agricultural systems.

Plant phenotyping relevance

綿花の脱葉率・吐絮率という植物状態を衛星画像から推定する超解像および機械学習手法を開発・比較しており、表現型取得・推定が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.