DeeplabV3をベースにした温室栽培における土壌モニタリングポイントの特定方法

Accurate positioning of water and fertilizer monitoring points in the soil of potted plants contributes to improving the level of precision agriculture management. This study proposes an automated extraction method for water and fertilizer detection points in blueberry potted plants in a greenhouse scenario. Based on the improved DeeplabV3+, a lightweight soil region segmentation network (LSRS-Net) is proposed. Firstly, the original DeeplabV3 + backbone was replaced by the lightweight MobileNetV2 to reduce computational complexity. Then, the Densely connected Atrous Spatial Pyramid Pooling (DenseASPP) module enhances multi-scale context feature fusion by embedding a Feature Pyramid Network (FPN) between the encoder and decoder to achieve cross-layer semantic fusion. In addition, the optimization algorithm based on Euclidean distance has improved the coordinates of the detection points. The experimental results show that, compared with the original model, the Mean Pixel Accuracy (MPA) and Mean Intersection over Union (MIoU) of the LSRS-Net model have increased by 2.98% and 3.84%, respectively, and the processing speed has reached 25.24 ms per frame. The extraction validity rate (EVR) exceeds 98.33%, which meets the requirements for fast and stable application in greenhouse cultivation and provides a viable solution for intelligent water and fertilizer management in agriculture.