The existing research on the planar form of traditional residences in the Hehuang region presents inconsistencies and insufficiencies in type division. Based on the review of existing research and the assessment of multi-source data, a new type division system is proposed and its applicability is validated by using convolutional neural networks (CNN) technology. The results show that: ① the overall recognition accuracy and the boundary definition of clusters in the new typological system are significantly superior to those of four existing systems; ② the inclusion of "Ⅱ-shaped" not only improves the recognition accuracy for "L-shaped" and reduces the misclassification rate for "Ⅰ-shaped" but also significantly enhances the semantic distinctiveness between the clusters of "L-shaped", "凹-shaped " and "Ⅰ-shaped". ③ the inclusion of "日-shaped" not only improves the recognition accuracy for "凹-shaped" forms and reduces the probability of misclassifying "凹-shaped" as "回-shaped", but also strengthens the semantic boundaries between the "凹-shaped " and " 回-shaped" clusters, enhancing the consistency of semantic features among the samples within each category. Overall, the newly established type system is more applicable to the division of types of traditional residential planar forms in the Hehuang region, providing theoretical support for the protection and inheritance of traditional villages and residences in the region, and offering case references for similar research in other areas.