Improved random patches and model transfer for deriving leaf mass per area across multispecies from spectral reflectance

Leaf mass per area (LMA) serves as a valuable metric within the field of agriculture, offering valuable insights into various aspects of leaf structure, including photosynthetic capacity, carbon assimilation, water use efficiency, and overall crop productivity. Machine learning, in combination with spectral reflectance analysis, has proven to be an highly advantageous approach for estimating LMA. The goal of this study is to develop machine learning models for LMA estimation across six multispecies datasets using leaf reflectance data at wavelengths from 400 to 2450 nm. The study evaluated and compared the performance of three regression models: partial least square regression (PLSR), eXtreme gradient boosting (XGBoost), and ensemble on random patches (RP). The results indicated that, in half of the validation cases, the PLSR model produced negative predictions, posing practical challenges for its application. In most instances, the RP model outperformed the XGBoost model, showcasing an average improvement of 104.27 % in terms of the coefficient of determination (R2). To further improve the estimation, transferring samples from the target dataset to the source dataset was implemented to enhance model performance across multispecies datasets. Meanwhile, an improved RP method specifically designed to sample transfer was proposed. The results showed a gradual improvement in LMA estimation as the number of transferred samples increased. When 10 % of new samples were transferred, both the standard and improved RP methods demonstrated substantial enhancements, including an average increase of 42.92 % in R2 and a reduction of 64.91 % in mean absolute error. Notably, the improved RP method consistently exhibited superior overall accuracy compared to the standard RP during the samples transferred. The proposed method represents a promising solution to overcome the challenges associated with leaf trait variability and environmental conditions.