首页|Crop type classification in Southern Brazil: Integrating remote sensing, crop modeling and machine learning
Crop type classification in Southern Brazil: Integrating remote sensing, crop modeling and machine learning
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
? 2022 Elsevier B.V.Crop type mapping is essential for agricultural monitoring, but lack of terrestrial labels (herein termed as ground truth data) limit these models around the globe. In this study, we tested different methods for creating crop type maps for soybean (Glycine max L.) and corn (Zea mays L.), with aims of i) compare ground truth data related to crop modeling; ii) evaluate agricultural field masks generated by Environmental Rural Register, MapBiomas product, and random forest model; iii) evaluate models of a) unsupervised classification, b) supervised classification with ground truth data of 1 year, c) with ground truth data of 2 years, d) supervised classification with crop modeling only, e) with the combination of crop modeling and ground truth data of 1 year, and f) with ground truth data of 2 years; iv) test sample size for training the model utilizing ground truth data and crop modeling in transfer learning approach; and v) examine spatial remote sensing features to guide crop data collection. The APSIM-NG crop model was utilized for generate crop modeling simulations to compare with satellite images harmonic regression of ground truth data. We found similarity of harmonic regression coefficients derived from crop modeling and satellite imagery of Sentinel-2 of the labeled ground truth data for both soybean and corn crops. Agricultural masks showed efficiency for crop area estimation for soybean, with highest accuracy with random forest model. Crop modeling aggregated with growing season data as input for supervised learning presented the greater model performance with overall accuracy of 0.94. Crop area prediction was most accurate for soybean [R2 = 0.93, and mean absolute error (MAE) = 2,052 ha] for the model with the combination of crop modeling and ground truth data of 2 years, and least for corn, [R2 = 0.18, and MAE = 1,146 ha] when only using crop modeling. Model performance was influenced by sample size, with greater accuracy (0.93) for aggregating crop modeling (150 samples) and year data (100 samples). In addition, considering spatial field data variability as model input increased overall accuracy from 0.84 to 0.93, with higher impact when only ground truth data was utilized for input in the model. Our results suggest that these methods offer options for crop type classification when less adequate ground truth data is available and with unsupervised learning. On the other hand, supervised learning utilizing crop modeling with presence of field data did improve model performance overall.
NSTL
Elsevier
