本研究旨在比較一個集成模型基線，該基線將來自每個作物的一個類別的子模型與使用多類作物訓練的單一模型相結合。數據集分為五輪訓練、驗證和測試。 Cohen-Kappa 公式用於找出基本事實，並根據有關背景、大米和捲心菜的基於包裹的信息進行預測。第一個實驗是比較沒有調整減重函數的集成模型和多作物單模型。結果是未調整重量損失函數的集成模型的平均 kappa (0.917) 高於多作物單模型的平均 kappa (0.897)。第二個實驗是比較沒有調整權重損失函數和調整權重損失函數的集成模型。結果是調整權重損失函數的集成模型的平均 kappa (0.953) 高於沒有調整權重損失函數的值。與多作物單模型的 kappa 相比，具有調整權重損失函數的 ensemble 模型的 kappa 有顯著差異。此比較的 Wilcoxon 有符號秩 p 值為 0.035。結論是，在本研究中，具有調整權重損失函數的集成模型比多作物單模型具有更好的性能。這意味著集成模型改進了多作物單模型。;The Convolutional Neural Networks (CNN) can be used in digital image processing, such as semantic segmentation. Using this model, an aerial image can be segmented to identify the crop in the parcel field. One of the CNN architectures that can perform semantic segmentation for aerial images is UNet-VGG16. This architecture is used in the architecture NCU RSS 1.3 which can segment rice in aerial images as in previous work of National Central University (NCU) last year. Furthermore, currently there is a cabbage dataset for another crop from the Council of Agriculture (COA) Executive Yuan that is ready for next NCU’s work.
The NCU RSS1.3 can be used in training models that can predict multi-crops of rice or cabbage in parcels of aerial images. However, the ensemble approach is also able to predict the crop in the aerial image because new crop datasets may be systematically added. The ensemble model will integrate the prediction outcomes from each crop′s sub-models. If the new crop dataset is ready, the sub-model of the previous dataset may not need to train in the ensemble model.
This study aims to compare an ensemble model baseline that combines sub-models from one class from each crop with the single model trained with multi-class crops. The dataset is split into train, validation, and test in five rounds. The Cohen-Kappa formula is used to figure out the ground truth and make predictions based on parcel-based information about the background, rice, and cabbage. The first experiment is comparing the ensemble model without tuning weight loss function and the multi-crops single model. The result is the average kappa of the ensemble model without tuning weight loss function (0.917) higher than the average kappa of the multi-crops single model (0.897). The second experiment is comparing the ensemble model without tuning weight loss function and with tuning weight loss function. The result is the average kappa of the ensemble model with tuning weight loss function (0.953) higher than without tuning weight loss function. The kappa of the ensemble model with tuning weight loss function is significantly different compared to the kappa of the multi-crops single model. The Wilcoxon Signed-Rank p-value of this comparison is 0.035. The conclusion is that the ensemble model with tuning weight loss function has better performance than the multi-crops single model in this research. It means that the ensemble model improves the multi-crops single model.
