Two-dimensional　(2D)　ferroelectric　field-effect　transistors　(Fe-FETs)　based　on　p-n　junctions　are　the　basic　units　of　future　neuromorphic　hardware.　The　In2Se3　semiconductor　with　ferroelectric,　photoelectric,　and　phase　transition　properties　possesses　great　application　potential　for　in-sensor　computing,　but　its　ferroelectric　p-n　junction　(FePNJ)　is　not　well　investigated.　Here,　we　present　an　optoelectronic　synapse　made　of　uniformly　full-coverage　alpha-In2Se3/WSe2　FePNJ,　achieving　ultralow-power　classification　recognition　and　multiscale　signal　processing.　Using　chemical　vapor　deposition　(CVD),　we　can　obtain　beta　＇-In2Se3/WSe2　subferroelectric　p-n　junctions　by　direct　growth　on　SiO2/Si　substrate　and　alpha-In2Se3/WSe2　FePNJ　by　phase　transition.　Modulated　by　the　synergistic　effect　of　the　polarization　electric　field　and　the　built-in　electric　field,　the　FePNJ　exhibits　significantly　enhanced　and　highly　tunable　synaptic　effects　(memory　retention　＞2500　s　and　＞8　multilevel　current　states　under　single　optical/electrical　pulses),　along　with　power　consumption　down　to　atto-joule　levels.　Utilizing　these　photoelectric　properties,　we　constructed　an　all-ferroelectric　in-sensor　reservoir　computing　system,　comprising　both　reservoir　and　readout　networks,　achieving　ultralow-power　handwritten　digit　recognition.　We　also　created　a　multiscale　reservoir　computing　system　through　the　gate-voltage-modulated　relaxation　time　scale　of　the　FePNJ,　which　can　efficiently　detect　motions　in　the　1　to　100　km　h(-1)　speed　range.