| This article has been：Browse 168Times Download 50Times |
scan it! |
|
|
| DOI：10.13522/j.cnki.ggps.2026015 |
|
| Estimating above-ground biomass of cotton using UAV remote sensing and interpretable machine learning with post-hoc feature selection |
|
Han Hongliang, Shi Haolei, Cao Hongxia, Wang Xuemei
|
|
1. School of Mechanical and Electrical Engineering, Shaanxi A&F Technology University, Yangling 712100, China;
2. Key Laboratory of Agricultural Soil and Water Engineering in Arid Areas, Ministry of Education, Northwest A&F University, Yangling 712100, China; 3. School of Water Conservancy Engineering, Shaanxi A&F Technology University, Yangling 712100, China
|
| Abstract: |
| 【Objective】Above-ground biomass (AGB) is an important indicator for evaluating healthy crop growth, and its traditional measurement was destructive and labour-intensive. The development of unmanned aerial vehicle (UAV) remote sensing and machine learning (ML) has enabled quick and accurate estimation of AGB. This study proposes interpretable ML models for cotton AGB estimation using UAV-acquired multispectral and RGB data.【Method】The field experiment was conducted from 2021 to 2022 during the cotton growing season. UAV equipped with multispectral and RGB sensors captured canopy imagery of the cotton field during the bud and flowering stages, from which a series of AGB-correlated colour and vegetation indices were extracted. Three ML algorithms, including Extra Trees, k-nearest neighbours (KNN), and support vector machine (SVM), were used to build AGB estimation models based on full-feature datasets. Shapley Additive Explanations (SHAP) were used to interpret the internal mechanisms of the constructed ML models. We proposed a SHAP-based post-hoc feature selection method and compared its performance with traditional priori feature selection approaches. Datasets from 2021 and 2022, in their independence or combination, were divided into training and test sets at a 7∶3 ratio. Ten-fold cross-validation was implemented during model training to alleviate overfitting and improve model robustness.【Result】Among all compared models, the Extra Trees algorithm performed optimally, with a coefficient of determination (R2) of 0.84, mean absolute error (MAE) of 1.22 t/hm2 and root mean square error (RMSE) of 1.48 t/hm2 for the test set. The combined 2021–2022 dataset required more feature indices to achieve optimal estimation accuracy than single-year dataset. SHAP interpretation effectively clarified the complex relationships between feature variables, AGB values, and ML predictions; model estimation accuracy depended on both data quality and the matching degree between feature selection schemes and algorithm characteristics. The vegetation index (NDVI) and colour indices (RGRI and EXG) were the dominant predictive features in all models. All three algorithms achieved reliable AGB estimation yet exhibited distinct feature utilization preferences. The Extra Trees model fully exploited vegetation index information, while KNN and SVM models relied more on colour indices. Compared with traditional priori feature selection methods, the proposed method effectively reduced the number of optimal feature variables and significantly improved estimation accuracy, with the SVM algorithm requiring the fewest input features. The optimal model varied with dataset types: the KNN model achieved the highest accuracy for the 2021 dataset (R2=0.85, MAE=1.07 t/hm2, RMSE=1.22 t/hm2); the Extra Trees model performed best for the 2022 dataset (R2=0.85, MAE=1.25 t/hm2, RMSE=1.63 t/hm2); the SVM model yielded the optimal performance for the combined 2021–2022 dataset (R2=0.80, MAE=1.48 t/hm2, RMSE=1.82 t/hm2).【Conclusion】Integrating the proposed post-hoc feature selection strategy, UAV remote sensing data, and interpretable ML algorithms can achieve accurate, non-destructive monitoring of cotton AGB in the field. |
| Key words: cotton; UAV remote sensing; machine learning; after-test; above-ground biomass; explainability |
|
|
