Photoacoustic imaging (PAI) can be used to monitor lesion formation during high-intensity focused ultrasound (HIFU) therapy because HIFU changes the optical absorption spectrum (OAS) of the tissue. However, in traditional PAI, the change could be too subtle to be observed either because the OAS does not change very significantly at the imaging wavelength or due to low signal-to-noise ratio in general. We propose a machine-learning-based method for lesion monitoring with multi-wavelength PAI (MWPAI), where PAI is repeated at a sequence of wavelengths and a stack of multi-wavelength photoacoustic (MWPA) images is acquired. Each pixel is represented by a vector and each element in the vector reflects the optical absorption at the corresponding wavelength. Based on the MWPA images, a classifier is trained to classify pixels into two categories: ablated and non-ablated. In our experiment, we create a lesion on a block of bovine tissue with a HIFU transducer, followed by MWPAI in the 690 nm to 950 nm wavelength range, with a step size of 5 nm. In the MWPA images, some of the ablated and non-ablated pixels are cropped and fed to a neural network (NN) as training examples. The NN is then applied to several groups of MWPA images and the results show that the lesions can be identified clearly. To apply MWPAI in/near real-time, sequential feature selection is performed and the number of wavelengths is decreased from 53 to 5 while retaining adequate performance. With a fast-switching tunable laser, the method can be implemented in/near real-time.