Research on Accurate Valuation of Farmland Assets and Plot Identification Model Based on UAV Multispectral Images

作者：Li Lin · 发表于：2025 Asia Conference on Energy Conversion Systems and Power Electronics (AECSPE) · 年份：2025 · DOI：10.1109/AECSPE66597.2025.00132

Aiming at the problems existing in current farmland asset valuation and land parcel identification, such as "vague land parcel boundary, large deviation in economic value evaluation and insufficient fusion of multi-source data", this study proposes a multi-modal framework integrating UAV five-band multi-spectral images, GIS and agricultural statistical information. Firstly, an improved U-Net++ land parcel recognition model is constructed, and CBAM attention mechanism and "spectrum-texture-topology" triple loss function are introduced. In the test of 1,258 land parcels in North China Plain, 86.9% mIoU and 0.83 boundary F1-score are achieved, which are 8.6% and 15.3% higher than the baseline model, and the recall rate of small land parcels is 82.7%. Secondly, the XGBoost+1D-CNN mixed estimation model is designed to extract 23-dimensional spectrum, texture, terrain and dynamic characteristics of growth period, and realize the joint prediction of yield and value at plot level. The estimated MAPE is reduced to 14.3%, which is significantly better than the single model. In the cross-validation of black soil region in Northeast China, the model still maintains an estimation error of 84.2% mIoU and 16.8%, which verifies its cross-regional generalization ability. This study provides a high-precision and transferable technical path for digital management of farmland assets.