Update app.py

app.py

@@ -853,36 +853,69 @@ async def predict_bloom(req: BloomPredictionRequest):
 
     # Decide status based on peak_score and bell shape
     bell_ok, bell_diag = is_bell_shaped(list(monthly_perc))
-    if peak_score < MIN_PEAK_FOR_BELL or not bell_ok or ambiguous:
+    mean_score = float(np.mean(monthly_scores))
+    median_score = float(np.median(monthly_scores))
+    near_peak_count = sum(1 for s in monthly_scores if s >= (peak_score * 0.9))
+    months_above_threshold = sum(1 for s in monthly_scores if s >= MIN_BLOOM_THRESHOLD)
+
+    print(f"DEBUG detection: peak_score={peak_score}, mean={mean_score:.2f}, median={median_score:.2f}, near_peak_count={near_peak_count}, months_above_{MIN_BLOOM_THRESHOLD}={months_above_threshold}, bell_ok={bell_ok}, bell_diag={bell_diag}")
+
+    # Heuristic decision:
+    # 1) If the absolute peak is below MIN_PEAK_FOR_BELL -> NO_BLOOM
+    if peak_score < MIN_PEAK_FOR_BELL:
         status = "NO_BLOOM"
         top_species = None
         bloom_window = None
         peak_month_out = None
         peak_prob_out = None
-    elif peak_score < MIN_BLOOM_THRESHOLD:
-        status = "LOW_CONFIDENCE"
-        top_species = None
-        bloom_window = [i+1 for i, p in enumerate(monthly_scores) if p > (MIN_BLOOM_THRESHOLD * 0.5)]
-        peak_month_out = peak_month
-        peak_prob_out = peak_score
     else:
-        status = "BLOOM_DETECTED"
-        bloom_window = [i+1 for i, p in enumerate(monthly_scores) if p > MIN_BLOOM_THRESHOLD]
-        peak_month_out = peak_month
-        peak_prob_out = peak_score
-
-        # species prediction when strong bloom signal
-        try:
-            peak_result = monthly_results[peak_idx]
-            if isinstance(peak_result, dict):
-                doy = peak_result.get("day_of_year")
-            else:
-                doy = date(year, peak_month, 15).timetuple().tm_yday
-            species_predictions = predict_species_by_elevation(elevation, doy=doy, top_k=TOP_K_SPECIES)
-            top_species = [SpeciesResult(name=sp, probability=round(prob * 100.0, 2)) for sp, prob in species_predictions]
-        except Exception as e:
-            print("❌ species prediction error at final stage:", e)
+        # 2) If peak is high enough, accept bloom even if not perfectly bell-shaped,
+        #    unless the season is extremely ambiguous (many months nearly at peak).
+        #    Use a dominance ratio: peak relative to mean/median.
+        dominance_ratio = (peak_score / (mean_score + 1e-9))
+        # Criteria to accept bloom:
+        accept_bloom = False
+        # a) clear bell-shaped + peak high -> accept
+        if bell_ok:
+            accept_bloom = True
+        # b) peak strongly dominates the rest -> accept
+        elif dominance_ratio >= 1.25 and near_peak_count <= 4:
+            accept_bloom = True
+        # c) several months above bloom threshold (broad season) but peak high -> accept but mark low confidence
+        elif months_above_threshold >= 3 and peak_score >= (MIN_BLOOM_THRESHOLD + 10):
+            accept_bloom = True
+
+        if not accept_bloom:
+            # Ambiguous/flat high-months case -> LOW_CONFIDENCE
+            status = "LOW_CONFIDENCE"
             top_species = None
+            bloom_window = [i+1 for i, s in enumerate(monthly_scores) if s > (MIN_BLOOM_THRESHOLD * 0.5)]
+            peak_month_out = peak_month
+            peak_prob_out = peak_score
+        else:
+            # Strong enough to declare bloom
+            # If the peak is only moderately high, use LOW_CONFIDENCE; otherwise BLOOM_DETECTED
+            if peak_score < (MIN_BLOOM_THRESHOLD + 5):
+                status = "LOW_CONFIDENCE"
+            else:
+                status = "BLOOM_DETECTED"
+
+            bloom_window = [i+1 for i, p in enumerate(monthly_scores) if p > MIN_BLOOM_THRESHOLD]
+            peak_month_out = peak_month
+            peak_prob_out = peak_score
+
+            # species prediction when we accept bloom
+            try:
+                peak_result = monthly_results[peak_idx]
+                if isinstance(peak_result, dict):
+                    doy = peak_result.get("day_of_year")
+                else:
+                    doy = date(year, peak_month, 15).timetuple().tm_yday
+                species_predictions = predict_species_by_elevation(elevation, doy=doy, top_k=TOP_K_SPECIES)
+                top_species = [SpeciesResult(name=sp, probability=round(prob * 100.0, 2)) for sp, prob in species_predictions]
+            except Exception as e:
+                print(f"❌ species prediction error at final stage: {e}")
+                top_species = None
 
     processing_time = round(time.time() - start_time, 2)