Google street view & deep learning: a new approach for ground truthing in crop mapping

Ground reference is the essential prerequisite for supervised crop mapping. However, conventional ground truthing involves extensive field surveys and post processes which is time-, labor- consuming and costly. Lacking a low-cost and efficient referencing method pervasively results in limited ground reference and impedes crop mapping but still attracted few attentions. In this study, we applied a convolutional neural network (CNN) model to explore the efficacy of generating ground truth from Google street view (GSV) images automatically in two distinct farming regions (center Illinois & southern California). We demonstrated the reliability and usability of the new ground reference further by performing a pixel based crop mapping. The same CNN model structure was employed by using various combination of vegetation indices as the model input. Results were evaluated with the USDA Crop Data Layer (CDL) products. From 8514 GSV images, the CNN model successfully screened out 2645 targeting crop images. The images were well classified into alfalfa, almond, corn, cotton, grape, soybean, pistachio and others. Overall GSV images classification accuracies reached 93% in California and 97% in Illinois. By shifting the geo coordinate of the images with the fixed empirical coefficients, we produced 8173 crop reference points inside parcels, 1764 in Illinois and 6409 in California. Evaluation of the new reference with CDL products showed satisfactory coherence with 81 to 97% agreement. Furthermore, the CNN based mapping also well captured the general pattern of crop type distributions. The overall differences between CDL products and our mapping results were 4% in California and 5% in Illinois. By embracing the power of deep learning and GSV images, we provided a new alternative for ground referencing and mapping crop types in an efficient and cost-effective way.