current pipeline

classification/classification_predict.py

@@ -82,7 +82,7 @@ def classify(img_path):
     # Initialize your custom model
     model = CustomNet(num_ftrs, num_classes)
     # Load the trained model weights
-    model.load_state_dict(torch.load('./fine_tuned:plant_classifier.pth'))
+    model.load_state_dict(torch.load('./classification/fine_tuned_plant_classifier.pth'))
 
     # Predict the class probabilities
     class_probabilities = predict_single_image(image_path, model)

detectree2model/predictions/predict.py

@@ -53,15 +53,17 @@ def predict(tile_path, overlap_threshold, confidence_threshold, simplify_value,
     download_file(url=url, local_filename=trained_model)
 
     cfg = setup_cfg(update_model=trained_model)
-    #cfg.MODEL.DEVICE = "cpu"
+
+    # hash the following line if you have gpu support
+    cfg.MODEL.DEVICE = "cpu"
     predict_on_data(tile_path, predictor=DefaultPredictor(cfg))
 
     project_to_geojson(tile_path, tile_path + "predictions/", tile_path + "predictions_geo/")
     crowns = stitch_crowns(tile_path + "predictions_geo/", 1)
     clean = clean_crowns(crowns, overlap_threshold, confidence=confidence_threshold)
     clean = clean.set_geometry(clean.simplify(simplify_value))
-    clean.to_file(store_path + "predicted_delineations.geojson")
+    clean.to_file(store_path + "detectree2_delin.geojson")
 
-def run_detectree2(tif_input_path, tile_width=20, tile_height=20, tile_buffer=20, overlap_threshold=0.35, confidence_threshold=0.2, simplify_value=0.2, store_path='./train_outputs/'):
+def run_detectree2(tif_input_path, store_path, tile_width=20, tile_height=20, tile_buffer=20, overlap_threshold=0.35, confidence_threshold=0.2, simplify_value=0.2): 
     tile_path = create_tiles(input_path=tif_input_path, tile_width=tile_width, tile_height=tile_height, tile_buffer=tile_buffer)
-    predict(tile_path=tile_path, overlap_threshold=overlap_threshold, confidence_threshold=confidence_threshold, simplify_value=simplify_value, store_path=store_path)
+    predict(tile_path=tile_path, overlap_threshold=overlap_threshold, confidence_threshold=confidence_threshold, simplify_value=simplify_value, store_path=store_path)

main.py

@@ -1,25 +1,77 @@
 from detectree2model.predictions.predict import run_detectree2
-from polygons_processing.postpprocess_detectree2 import postprocess, export_df_as_geojson
+from polygons_processing.postpprocess_detectree2 import postprocess
 from generate_tree_images.generate_tree_images import generate_tree_images
 from classification.classification_predict import classify
 import os
+import json
 
-if __name__=="__main__":
-    tif_input = ""
-    detectree2_output = run_detectree2(tif_input)
+def row_to_feature(row):
+    feature = {
+        "id": row["id"],
+        "type": "Feature",
+        "properties": {"Confidence_score": row["Confidence_score"]},
+        "geometry": {"type": "Polygon", "coordinates": [row["coordinates"]]},
+        "species": row['species']
+    }
+    return feature
 
-    processed_output_df = postprocess(detectree2_output)
-    processed_geojson = '../polygons_processing/postprocessed_predictions'
+def export_geojson(df, filename):
+    features = [row_to_feature(row) for idx, row in df.iterrows()]
 
-    generate_tree_images(processed_geojson, tif_input)
-    output_folder = '../generate_tree_images/tree_images'
+    feature_collection = {
+        "type": "FeatureCollection",
+        "crs": {"type": "name", "properties": {"name": "urn:ogc:def:crs:EPSG::32720"}},
+        "features": features,
+    }
 
-    for file_name in os.listdir(output_folder):
-        file_path = os.path.join(output_folder, file_name)
-        probs = classify(file_path)
-        #add probs to df and then convert to geojson again for final output
-        processed_output_df['class_probs'] = probs
-    
-    final_output_path = 'result.geojson'
-    export_df_as_geojson(processed_output_df, final_output_path)
+    output_geojson = json.dumps(feature_collection)
+
+    with open(f"{filename}.geojson", "w") as f:
+        f.write(output_geojson)
+
+    print(f"GeoJSON data exported to '{filename}.geojson' file.")
+
+"""
+tif_input: the file containing a tif that we are analyzing
+
+tif_file_name: the file name of the tif input. tif_input is the folder in which the tif file lies 
+(detectree2 works with that) but generate_tree_images requires path including the file hence the file name is needed
+
+output_directory: the directory were all in-between and final files are stored
+
+generate_tree_images stores the cutout tree images in a separate folder
+"""
+
+tif_input = "/Users/jonathanseele/ETH/Hackathons/EcoHackathon/WeCanopy/test/"
+tif_file_name = "TreeCrownVectorDataset_761588_9673769_20_20_32720"
+
+current_directory = os.getcwd()
+output_directory = os.path.join(current_directory, "outputs")
+if not os.path.exists(output_directory):
+    os.makedirs(output_directory)
+
+run_detectree2(tif_input, store_path=output_directory)
+
+processed_output_df = postprocess(output_directory + '/detectree2_delin.geojson')
+processed_geojson = output_directory + '/processed_delin.geojson'
+
+generate_tree_images(processed_geojson, tif_input)
+output_folder = './tree_images'
+
+all_top_3_list = []  # Initialize an empty list to accumulate all top_3 lists
+
+for file_name in os.listdir(output_folder):
+    file_path = os.path.join(output_folder, file_name)
+    probs = classify(file_path)
+    top_3 = probs.head(3)
+    top_3_list = [[cls, prob] for cls, prob in top_3.items()]
     
+    # Accumulate the top_3_list for each file
+    all_top_3_list.append(top_3_list)
+
+# Assign the accumulated top_3_list to the 'species' column of the dataframe
+processed_output_df['species'] = all_top_3_list
+
+final_output_path = 'result'
+export_geojson(processed_output_df, final_output_path)
+

polygons_processing/postpprocess_detectree2.py

@@ -351,6 +351,6 @@ def postprocess(prediction_geojson_path):
 
     df_res = process([df])
 
-    export_df_as_geojson(df=df_res, filename="postprocessed_predictions")
+    export_df_as_geojson(df=df_res, filename="processed_delin")
 
     return df_res