Integration of multimodal data for large-scale rapid agricultural land evaluation using machine learning and deep learning approaches

• DOI: 10.1016/j.geoderma.2023.116696
• Publication Date: 2023-10-25T09:39:10Z
• Authors: Liangdan Li; Luo Liu; Yiping Peng; Yingyue Su; Yueming Hu; Runyan Zou
• Keywords: Agricultural land

Rapid and accurate agricultural land evaluation provides essential guidance for the supervision allocation of resources; it also helps to ensure food security. Previous work has mainly evaluated quality at county level by using field sampling data based on a factor approach. However, is difficult achieve uniform, large-scale via conventional approaches because its spatial heterogeneity, as well large temporal economic costs associated with acquisition. In this study, we integrated publicly available multimodal (i.e., satellite remote sensing, environmental, socioeconomic data) into Google Earth Engine (GEE) platform, selected best indicators from each modality geodetector, basis which different combinations input models were designed. And then developed machine learning (random forest, RF) deep (deep neural network, DNN) evaluate in paddy dry systems 2013 throughout Guangdong Province, China. The results showed that performance our combination variables decreased following order: > bimodal unimodal. With combination, RF model (R2 = 0.91, RMSE 97.56, CCC 0.95) outperformed DNN 0.89, 108.72, 0.94) terms predicting field. 0.90, 104.27, 0.86, 124.38, 0.93) land. estimates obtained more than greater homogeneity fields. This research proposed simple, low-cost rapid provincial scale data, can help control grade multiple scales.