The accurate, efficient and nondestructive detection of chlorophyll content in rice leaves using spectral information is of practical importance for diagnosing and optimizing nitrogen nutrition in rice leaves, developing and optimizing nitrogen fertilization systems in rice fields, and monitoring and evaluating rice pests and diseases. This paper addresses the problem of poor model accuracy and stability when machine learning models are used solely to invert the chlorophyll content of rice leaves. Moreover, takes Northeast japonica rice Jijing88 as the research object, obtains leaf phenotypic hyperspectral data and relative chlorophyll content at key fertility stages such as tillering through grid tests, select the kernel limit learning machine (Kernel function extreme learning machine, KELM) in machine learning as the base modeling model, and proposes a new idea of selecting preprocessing methods based on the base KELM modeling effect first, and then optimizing the KELM training process corresponding to the selected preprocessing combination using a bionic optimization algorithm to improve the model prediction accuracy. First, this paper investigates the preprocessing methods of spectral data, and a total of 72 preprocessing combinations are obtained by combining all four types of preprocessing methods. The sequential projection algorithm (Successive Projections Algorithm, SPA) is used to select the characteristic bands for input into the KELM model to filter the better preprocessing combinations. Based on the modeling effect, the test set’s coefficient of determination (R2p) corresponding to KELM for the pretreatment combinations CWT+MMS, CWT+MSC+SG+SS, and CWT+SS was higher, 0.850, 0.835, and 0.828, respectively. Secondly, to make the KELM model perform optimally while ensuring stability and generalization. In this paper, the Harris Hawk Optimization Algorithm (Harris Hawks Optimizer, HHO) is introduced to automatically and optimally adjust the parameters of the above three KELM models by simulating the cooperative behavior and chasing strategy of the hawks during predation, resulting in the HHO-KELM models with R2p of 0.957, 0.867 and 0.858, respectively, and a maximum of 10.7% effectively improves the model accuracy. The feasibility of the HHO algorithm to optimize the machine learning model to invert the chlorophyll content of rice leaves was demonstrated, which provides a strong reference and reference for the determination and assessment of chlorophyll content in northeastern japonica rice.