Estimating fuel load of Pinus massoniana coniferous layer beds using three machine learning algorithms integrated with image features

This study combined artificial intelligence and image recognition technologies to construct a machine learning predictive model based on image features was constructed to achieve rapid, objective, and accurate estimation of surface fuel loading in coniferous forests, providing data support for refined forest fire prevention, forecasting, and related endeavors.

Taking the needles of typical Pinus massoniana forests in Guizhou Province were selected as the research object. The actual variation range of fuel loading in the needle litter layer was determined through setting up sample plots and conducting random quadrat surveys. In the laboratory, needle litter layers with different fuel loading (30 cm×30 cm) were constructed, and each fuel loading was photographed three times repeatedly, resulting in a total of 150 vertically oriented images. Image features were extracted using OpenCV-Python. Following Z-score standardization and principal component analysis (PCA) for feature selection, three machine learning methods were employed to construct prediction models for needle litter layer loading, and the model performance was evaluated.

1) Field measurements revealed that the fuel loading of the needle litter layer ranged from 3.9-16.5 t·hm^-2. Among the shape features extracted from images of litter layers with different fuel loading, the maximum perimeter value reached 600 223.95, while the edge density in edge features was the smallest, with a value of 4.11; 2) Severe multicollinearity (VIF>10) was observed among all eight image feature values; 3) The first three principal component scores (PC1, PC2, and PC3), extracted via PCA, were used as new independent variables. The fuel loading showed statistically highly significant linear correlations with PC1, PC2, and PC3, indicating their suitability for model construction; 4) Machine learning prediction models were constructed using the training set and validated with the testing set. The K-nearest neighbor (KNN) model achieved the best prediction performance, with a mean relative error (MRE) of 18.64%, mean absolute error (MAE) of 2.29 t·hm^-2, root mean square error (RMSE) of 3.31 t·hm^-2, R-squared (R²) of 0.79. Its low Jensen-Shannon divergence (JSD) value of 0.005 indicated highly similar distributions between predicted and true values. The random forest regression (RFR) model performed suboptimally (MRE=39.03%, R²=0.57), while the multiple linear regression (MLR) model yielded the worst prediction results.

The machine learning model constructed based on image features proved feasible for estimating the fuel loading of the needle litter layer. It overcame the drawbacks of traditional measurement methods, such as being time-consuming, subjective, and low in accuracy, achieving rapid and objective estimation of fuel loading. This approach provided a novel methodology for research on forest surface fuel loading and held significant implications for forest fire forecasting and scientific management.