Integrating Random Forest With Boundary Enhancement for Mapping Crop Planting Structure at the Parcel Level From Remote Sensing Images

Accurately and efficiently obtaining crop planting structure information is critical for precision agriculture. However, the current methods for mapping crop planting structure primarily use image pixels as the classification units, easily leading to blurred and fragmented boundaries and the salt-and-pepper effect, which significantly limit the accuracy and reliability of the results. To address this challenge, we propose a novel framework for mapping crop planting structure, consisting of three key components: 1) farmland parcel extraction; 2) crop classification feature extraction; and 3) crop classification. First, a boundary-enhanced deep-learning model is introduced for farmland parcel extraction (FPENet) from Gaofen-2 data, based on the U-Net model, to accurately obtain farmland parcel data. Subsequently, crop classification features are extracted at the parcel level from both Sentinel-2 and Landsat 8 data. After selecting the optimal feature combination, crop classification is performed using the random forest model to map precise crop planting structure. The proposed framework was evaluated in Dangyang County, Hubei province, China, where it showed a superior performance in mapping crop planting structure. The FPENet model achieved an overall accuracy and F1-score exceeding 92.5%, enabling complete and accurate extraction of farmland parcels. Comparative experiments with different convolutional neural networks further highlighted FPENet's exceptional capability. Furthermore, with the optimal feature combination, the classification accuracy for rice, corn, and wheat exceeded 94.5%, with spectral bands and vegetation indices being the key contributors to crop classification. In addition, comparisons with other methods further validated the effectiveness of this framework in mapping crop planting structure.