@article{oai:ipsj.ixsq.nii.ac.jp:02002190,
 author = {眞田,慎 and 黒田,剛士 and ダマリス,トリフェナ and 猿田,悠 and 峰野,博史 and 難波,直樹 and 内海,智仁 and Makoto Sanada and Tsuyoshi Kuroda and Damaris Trifena and Yu Saruta and Hiroshi Mineno and Naoki Namba and Tomoyoshi Utsumi},
 issue = {2},
 journal = {情報処理学会論文誌コンシューマ・デバイス＆システム（CDS）},
 month = {May},
 note = {ワインブドウ栽培において収量予測は，生産性向上や適切な雇用管理に重要である．古典的な収量予測手法では樹木あたりの平均房数と平均房重量を農家が手作業で計測するため，労力や時間などのコストが高く，収穫期前に房を採取する必要があった．近年は画像ベースの深層学習モデルを用いた低コストで採取を要さない手法が提案されているが，葉による房のオクルージョンが予測精度の向上を妨げている．本論文では葉が少ないブドウ樹木の生育初期に着目し，オクルージョンの影響を抑えた収穫房数予測に対する芽や若枝カウントの有効性を評価する．さらに，平均房重量の予測に対する房の推定インスタンスセグメンテーション結果の有用性を示し，早期収量予測システムの実現可能性を提示する．本実験ではブドウの生育期を新芽/若枝/花序/房に分け，各生育期で新芽などの対象カウントに基づく収穫数予測の精度を比較した．予測と目視のカウント誤差は新芽期で最も小さく2.4%に収まり，オクルージョンの影響を受けない収穫房数が収穫期6カ月前から得られることが示された．また，房のインスタンスセグメンテーション結果から算出した平均房体積の時系列変化が，一般的なブドウ成長曲線と類似することが確認できた．深層学習モデルを用いた早期収穫数予測と房体積の推定によって正確な収量予測が実現可能であることが示せた．, Early yield prediction is necessary to improve productivity and employment management in precision viticulture. In the conventional method, farmers manually measure the average bunch number per vine and the average bunch weight for yield prediction. This method is destructive and costly in terms of labor and time. Recently, low-cost, non-destructive methods using image-based deep learning models have been researched. However, it is difficult to obtain sufficiently accurate predictions due to the occlusion of bunches by leaves. In this study, we focus on the vine's early growth period that does not occur occlusion by leaves, and evaluate that the number of harvest bunches can be predicted based on the number of buds or shoots counted with a deep learning model. We also discuss the usefulness of the estimated instance segmentation results of bunches for predicting average bunch weight, and demonstrate the feasibility of an early yield prediction system. In this experiment, the grape growth period is divided into Bud/Shoot/Inflorescence/Bunch, and the number of harvest bunches is predicted in each period from the number of objects, e.g., buds. The error between the predicted harvest number and visual counting was smallest (about 2%) in the Bud period. This result shows that counting at the early growth period is useful for harvest number prediction. The time series change of estimated bunch volume based on instance segmentation was comparable to the widely accepted berry growth curve. These results show that yield prediction with high accuracy is possible by the early harvest number prediction and the bunch volume estimation based on instance segmentation.},
 pages = {11--20},
 title = {ワインブドウの収穫期6カ月前でも得られる早期収量予測の指標について},
 volume = {15},
 year = {2025}
}