The PhenoCam network is a global array of near-surface cameras used to track vegetation processes. With hundreds of cameras taking up to 48 photos a day this data stream has provided novel insight into many biological processes, yet PhenoCams in agricultural research have been underutilized. Here we use images from LTAR agricultural sites in the PhenoCam network to build a classification model for identifying cropland attributes such as crop type and phenological states. A variety of crops are used to generalize the states into 21 classes across 3 mutually exclusive categories, ranging from emergence to senescence. Other factors such as crop type and flooded or snow covered fields are also included. Deep learning models designed for image classification do not have a temporal component, so we use a hidden markov model in the classification post-processing to account for the temporal autocorrelation of daily camera time series. Initial classification F1 scores, a 0-1 accuracy metric, ranged from 0.83-0.85 with held out validation data, and were increased to 0.86-0.87 after the post-processing steps. Results show the feasibility of a daily, local scale dataset of field states and phenological stages with potential to be applied to other near-surface cameras for agroecosystem monitoring.