BLOG/UPDATES

ResUnet support added

ResUnet architecture added as an option to deepTerra. ResUNet is a convolutional neural network that combines U-Net’s encoder–decoder structure with ResNet-style skip connections to improve feature learning and segmentation accuracy, especially on complex images. Applied it to the Cyprus garbage detection domain, where is showed promising results.

Quarries - dataset available

The training and test data used for the Cyprus Quarries investigation is available. Refer to the resources page.

Cyprus Quarries

A new project has been completed to chart the number and location of Cyprus quarries. A supervised learning model was trained and tested against more than 300 verified quarry image patches. The google earth API was used to download a fresh set of satellite images covering almost the entire island. This was divided into more than 185,000 image patches, each representing an area of around 220x220m. When the model was run against this dataset, around 2000 quarry locations were identified, with an accuracy of more than 90%. Full details can be found here.

Garbage Detection Kouris Riverbed

A new (smallish) garbage analysis has been made, this time for the area in Cyprus immediately south of the Kouris reservoir. This covers 6 km^2 following the line of the (usually) dried up river bed, up to the Erimi/Episkopi border. This includes reference to more than 900 suspected garbage dumps, of varying sizes. This can be found at: Garbage map summary. Click one of the colored squares to drill down further.

Garbage detection updates

I've made a few updates to my AI based garbage detection system.

1. I've a model that better filters out "trivial" garbage, hopefully focusing on more significant dumps. It's still far from perfect, and is still a work in progress.
2. I've created a more device/memory friendly overview of the area covered (roughly 200 sq km covering the Pitsilia region of Cyprus) which shows a checkerboard of 280 approx 200x200m patches. Clicking on one of these pops up a small dialog which details how many garbage sites it has found within that patch, together with a link that pops up a more detailed window showing where it thinks that garbage is. This can be found at: Garbage map summary

Multi-classification via blood cell recognition

Until now, deepTerra has been used exclusively with binary classification datasets. It has now been extended to support multi-class classification problems. As a case study, it was applied to a Peripheral Blood Cell (PBC) dataset containing 17,092 images across eight distinct cell types: neutrophils, eosinophils, basophils, lymphocytes, monocytes, immature granulocytes (including promyelocytes, myelocytes, and metamyelocytes), erythroblasts, and platelets (thrombocytes). The original dataset can be found here. An example image from each category is shown below.

The dataset was divided into a training and test dataset (50/50 split) and the standard ResNet152 architecture pretrained on ImageNet then fine-tuned using the PBC training dataset. To improvde performance, the final classification layer was substitued with a Support Vector Machine (SVM). The result was tested against the test dataset with the following results:

Class	Precision	Recall	F1-Score	Support
basophil	0.9015	0.9770	0.9377	609
eosinophil	1.0000	0.8999	0.9473	1559
erythroblast	0.9878	0.9369	0.9616	776
immature granulocytes	0.8441	0.9537	0.8956	1448
lymphocyte	0.9827	0.9357	0.9586	607
monocyte	0.9361	0.8662	0.8998	710
neutrophil	0.9296	0.9514	0.9403	1665
platelet	0.9991	0.9991	0.9991	1174
Overall Accuracy				0.9413
Macro-average F1 Score				0.9425
MCC				0.9319

An accuracy of > 94% is pretty good (certainly much better than I can do myself) and illustrates the power of these techniques.

Investigation into SVM on top of CNN

Various research papers have explored the use of hybrid CNN–SVM architectures for classification tasks. One such example is An Architecture Combining Convolutional Neural Network (CNN) and Support Vector Machine (SVM) for Image Classification.

Inspired by these findings, I experimented with this hybrid approach on one of my garbage detection datasets. I used the CNN's second-to-last layer to extract high-level feature embeddings from the images, replacing the final softmax classification layer with an SVM.

The results were promising. Without the SVM, the model achieved an accuracy of 0.75 and a Matthews Correlation Coefficient (MCC) of 0.52. With the SVM classifier, accuracy improved to 0.84 and MCC to 0.69.

Based on these results, I’ve added a new enhancement task to the deepTerra backlog to support SVM-backed classification pipelines.