Colibry Forest Sentinel

Ultra-low-power AI device that detects chainsaw sounds in forests to fight illegal logging.

IntermediateWork in progress213

Things used in this project

Hardware components

ASYGN ColibryNPU EvalboardXP

MIKROE LR IoT Click

Seeed Studio SenseCAP M2 Multi-Platform LoRaWAN Indoor Gateway(SX1302) - EU868

Software apps and online services

Helium LoRa Network

Optional

Story

Illegal logging is one of the greatest threats to forests worldwide. It not only drives large-scale deforestation but also affects biodiversity, the climate, and local communities that rely on forests for their livelihoods. According to the FAO, millions of hectares of forest are lost each year due to unregulated logging activities, often in remote areas where traditional human surveillance is nearly impossible.

Fighting this threat presents a major challenge: how can we effectively protect vast and isolated territories while minimizing costs and environmental impact? Conventional solutions, such as on-site patrols or video surveillance, face significant logistical and financial limitations.

In this context, our project is to protect forests by detecting illegal logging in real time. We designed a tiny, ultra-low-power AI classification algorithm capable of recognizing the sound of chainsaws in remote forest areas.

By embedding an intelligent sound recognition model directly into Colibry, an ultra low-power microcontroller, our device can listen to the forest continuously, detect suspicious noises, and trigger alerts – all without relying on heavy infrastructure or consuming much energy.

How the system works:

The system is built around a Colibry board equipped with an integrated microphone and a LoRa communication module.

Colibry board with LoRa module

The board runs a deployed model capable of detecting chainsaw sounds in forest environments. When such sounds are detected, the system sends an alert via LoRa to a connected gateway (using a community-driven LoRa network is also possible), enabling real-time notification of potential illegal logging activities.

LoRa Gateway

About the device:

Real time chainsaw detection and monitoring
Can be self-sufficient with solar energy
Operates fully offline
1mW data processing on Colibry

Monitoring

Model:

The model is based on a modified BlazeFace backbone, adapted to process spectrogram inputs.

After training, the model was quantized to reduce memory footprint and improve inference efficiency. Finally, it was converted into a bitstream representation, making it suitable for deployment on the Colibry chip.

To train the model we used the public chainsaw audio dataset from Rainforest Connection available here: https://huggingface.co/datasets/rfcx/frugalai

We achieved 93.9% accuracy on the test set

Inference requires 1634708 cycles

Bill of materials

Colibry: $11.00 (in volume)

MIKROE LR IoT Click: $54.00

SeeedStudio SenseCAP M2 Multi-Platform LoRaWAN Indoor Gateway(SX1302) - EU868: $139.00 (Optional)

#include "lriot_chainsaw.h"
#include "driver_colibry.h"
#include <stdio.h>
#include <stdbool.h>
#include <string.h>

#define AUDIO_BLOCK_SIZE 128
#define FFT_SIZE         256
#define NPU_THRESHOLD    0.5f

// --- Circular buffer for DMA I2S ---
float audio_circ_buffer[FFT_SIZE];
volatile int dma_write_index = 0;  // Updated by DMA ISR
volatile bool fft_ready = false;    // Set by DMA ISR when buffer is full

float fft_output[FFT_SIZE];

// --- DMA callback stub ---
// This function should be called by the DMA interrupt
void i2s_dma_half_full_callback(const float *dma_buffer, int len)
{
    // Copy new samples into the circular buffer
    for (int i = 0; i < len; i++) {
        audio_circ_buffer[dma_write_index++] = dma_buffer[i];
        if (dma_write_index >= FFT_SIZE) {
            dma_write_index = 0;
            fft_ready = true;  // Full buffer ready for FFT
        }
    }
}

// --- Main ---
int main(void)
{
    // --- Initialize peripherals ---
    i2s_init_dma(audio_circ_buffer, FFT_SIZE); // DMA-based I2S, double-buffering
    fft_init(FFT_SIZE);
    npu_init();
    if (npu_load_model("chainsaw.nnpa") != 0) {
        printf("[NPU] Failed to load model\n");
        return -1;
    }

    if (lriot_chainsaw_init() != 0) {
        printf("[LoRaWAN] Init failed\n");
        return -1;
    }

    printf("[System] DMA-based pipeline initialized\n");

    // --- Main loop ---
    while (1)
    {
        // Check if FFT buffer is ready (set by DMA ISR)
        if (fft_ready)
        {
            fft_ready = false;

            // 1. Compute FFT on the full buffer
            fft_compute(audio_circ_buffer, fft_output, FFT_SIZE);

            // 2. Run NPU detection
            float detection_score = npu_process(fft_output, FFT_SIZE);

            // 3. Fire-and-forget LoRaWAN send if detected
            if (detection_score > NPU_THRESHOLD) {
                int ret = lriot_chainsaw_send_hex();
                if (ret == 0)
                    printf("[LoRaWAN] Event sent\n");
                else
                    printf("[LoRaWAN] Send failed\n");
            }
        }
        Delay_ms(1);
    }

    return 0;
}

#include "lriot_chainsaw.h"
#include <stdio.h>

static lriot_t lriot;
static lriot_cfg_t cfg;

int lriot_chainsaw_init(void)
{
    lriot_cfg_setup(&cfg);
    LRIOT_MAP_MIKROBUS(cfg, MIKROBUS_1);

    if (lriot_init(&lriot, &cfg) == LRIOT_ERROR) {
        return -1;
    }

    lriot_reset(&lriot);
    Delay_ms(2000);

    lriot_cmd(&lriot, "mac set deveui 0011223344556677");
    lriot_cmd(&lriot, "mac set appeui 0011223344556677");
    lriot_cmd(&lriot, "mac set appkey 00112233445566778899AABBCCDDEEFF");

    if (lriot_cmd(&lriot, "mac join otaa") != LRIOT_OK) {
        return -1;
    }

    return 0;
}

int lriot_chainsaw_send_hex(void)
{
    const char *payload_hex = "636861696e736177"; // "chainsaw" in hex
    char cmd_buf[64];
    snprintf(cmd_buf, sizeof(cmd_buf), "mac tx uncnf 1 %s", payload_hex);

    if (lriot_cmd(&lriot, cmd_buf) != LRIOT_OK) {
        return -1;
    }

    return 0;
}