In Hong Kong, the natural terrain is susceptible to rain induced landslides. These landslides are usually of small-to-medium scale, involving the failure of soil within the top one to two meters of the surface mantle. A comprehensive historical landslide database and distribution of geological features are crucial for understanding the landslide susceptibility of natural terrain. The location of natural terrain landslides and other geological features are currently identified from aerial photograph interpretation (API) by experienced engineering geologists. With about 10,000 aerial photographs taken annually, there are strong initiatives to apply machine learning to facilitate the identification process. A method combining machine learning technology and image analysis methodology was developed to help automatically and objectively acquire the location and geometric information of landslides. The model was trained using geo-referenced aerial photographs together with manually mapped landslide boundaries within pilot study areas in Hong Kong. The trained model was then applied to extract landslide data from aerial photographs taken at other areas and time with promising results. Similar machine learning techniques can also be utilized to identify geological features, such as rock outcrops, from remote sensing imageries. Indeed, a territory-wide rock outcrop map for the natural terrain of Hong Kong has been produced using such approaches. The above applications can provide useful data on landslide susceptibility and facilitate the identification of vulnerable catchments for natural terrain hazard studies. This paper introduces the workflows and the architecture design of the neural networks applied. The extraction results, the applications of the techniques and the way forward are discussed.