Stream　-observing　cameras　have　recently　been　deployed　in　stream　systems　to　monitor　water　depth　dynamics.　However,　most　existing　image　-based　water　depth　monitoring　methods　require　additional　gauging　equipment,　extensive　manual　annotations,　or　complex　manual　calibration.　In　this　paper,　we　propose　the　hierarchical　model,　a　novel　multi　-modal　and　multi　-scale　deep　learning　framework　for　monitoring　water　depth　in　headwater　streams　with　only　a　field　camera　capable　of　night　vision　and　no　additional　equipment.　In　particular,　the　hierarchical　model　integrates　long-term　dynamic　patterns　extracted　from　large-scale　meteorological　data　with　short-term　dynamic　patterns　extracted　from　small-scale　stream　image　data　to　jointly　monitor　water　depth　at　a　fine　-level　temporal　resolution.　In　order　to　overcome　the　issue　of　limited　availability　of　images,　we　introduce　a　transfer　learning　strategy　and　incorporate　more　accurate　long-term　patterns　that　enable　the　hierarchical　model　to　perform　competitively　even　with　a　small　number　of　images.　We　evaluate　our　method　on　a　real　-world　headwater　stream　monitoring　dataset　from　the　West　Brook　study　area　in　western　Massachusetts,　United　States.　Our　extensive　experiments　demonstrate　that　the　hierarchical　model　outperforms　several　state-of-the-art　methods　for　water　depth　monitoring,　and　that　more　accurate　long-term　patterns　can　better　guide　the　monitoring　of　short-term　patterns　with　excellent　flexibility　and　less　computational　cost.　The　mean　absolute　error　of　our　hierarchical　model　achieves　a　remarkable　level　of　4　.　9　cm　at　the　study　site　with　0　.　89　m　average　water　depths,　and　only　12　.　5　cm　at　more　drastically　varied　site　with　3　.　95　m　average　depths.